def init_with_data(self, init_X, init_Y):
"""
Input parameters
----------
gp_params: Gaussian Process structure
x,y: # init data observations (in original scale)
"""
# Turn it into np array and store.
self.X_original = np.asarray(init_X)
temp_init_point = np.divide((init_X - self.bounds[:, 0]),
self.max_min_gap)
self.X_original = np.asarray(init_X)
self.X = np.asarray(temp_init_point)
self.Y_original = np.asarray(init_Y)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.NumPoints = np.append(self.NumPoints, len(init_Y))
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
def test_acquisition_functions(acquisition_function: acquisition_functions.AcquisitionFunction):
batch_size = 13
test_loader = torch.utils.data.DataLoader(
datasets.MNIST(
"../data",
train=False,
transform=transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]),
),
batch_size=batch_size,
shuffle=False,
)
bayesian_net = mnist_model.BayesianNet()
estimator = acquisition_function.create(bayesian_net, k=1)
estimator.eval()
scores = torch.tensor([])
num_iters = 5
for data, _ in itertools.islice(test_loader, num_iters):
output = estimator(data)
scores = torch.cat((scores, output), dim=0)
assert scores.shape == (batch_size * num_iters,)
def test_check_input_permutation(af_type: acquisition_functions.AcquisitionFunction):
if af_type == acquisition_functions.AcquisitionFunction.random:
return
batch_size = 12
test_data = torch.rand((batch_size, 10))
mixture_a = test_data[::2, :]
mixture_b = test_data[1::2, :]
mixture_c = test_data
class Forwarder(torch.nn.Module):
def forward(self, batch):
return batch
forwarder = Forwarder()
estimator = af_type.create(forwarder, k=1)
estimator.eval()
output_a = estimator(mixture_a)
output_b = estimator(mixture_b)
output_c = estimator(mixture_c)
torch.testing.assert_allclose(
torch.cat([output_a, output_b], dim=0), torch.cat([output_c[::2], output_c[1::2]], dim=0)
)
def expandBoundsDDB_FB(self):
"""
Description: Expands the search space with the full Bayesian
implementation of our DDB method
"""
print('Attempting to expand search space with DDB-FB method')
alpha=self.alpha
beta=self.beta
bound_samples=100 # Number of radius sample to fit the log-logistic distribution
# Find y^+ and x^+
ymax=np.max(self.Y)
# Generate test radii
max_loc=np.argmax(self.Y)
xmax=self.X[max_loc]
test_bound=np.zeros(self.scalebounds.shape)
bound_dist=np.zeros(bound_samples)
bound_center=xmax
test_bound[:,1]=bound_center+0.5
test_bound[:,0]=bound_center-0.5
max_radius=np.max(np.array([np.max(max_bound_size-test_bound[:,1]),np.max(test_bound[:,0])]))
step=max_radius/bound_samples
packing_number=np.zeros(bound_samples)
# Generate a Thompson sample maxima to estimate internal maxima
TS=AcquisitionFunction.ThompsonSampling(self.gp)
tsb_x,tsb_y=acq_max_global(TS, self.gp, bounds=self.scalebounds)
# Generate Gumbel samples to estimate the external maxima
for i in range(0,bound_samples):
bound_length=test_bound[:,1]-test_bound[:,0]
volume=np.power(max_bound_size,self.dim)-np.prod(bound_length)
packing_number[i]=round(volume/(5*self.gp.lengthscale))
mu=stats.norm.ppf(1.0-1.0/packing_number[i])
sigma=stats.norm.ppf(1.0-(1.0/packing_number[i])*np.exp(-1.0))-stats.norm.ppf(1.0-(1.0/(packing_number[i])))
bound_dist[i]=np.exp(-np.exp(-(-tsb_y-mu)/sigma))
test_bound[:,1]=test_bound[:,1]+step
test_bound[:,0]=test_bound[:,0]-step
bound_dist[np.isnan(bound_dist)]=1
# Fit the log-logistic paramaters to the Gumbel samples
xfit=np.arange(0,max_radius,max_radius/100)
popt,pcov=optimize.curve_fit(self.sufficientBoundPDF,xfit[0:100],bound_dist,bounds=np.array([[5,1.1],[20,5]]))
print("popt={}".format(popt))
b=ymax/popt[0]
a=popt[1]
print("b={}, ymax={}".format(b,ymax))
# Sample for the optimal radius
for d in range(0,self.dim):
gamma=np.random.gamma(shape=alpha,scale=1/beta,size=100)
loglog=stats.fisk.pdf(gamma,ymax/b,scale=a)
scaled_weights=loglog/np.sum(loglog)
multi=np.random.multinomial(1,scaled_weights)
r_index=np.argmax(multi)
print("Radius of {} selected".format(gamma[r_index]))
self.scalebounds[d,1]=xmax[d]+gamma[r_index]
self.scalebounds[d,0]=xmax[d]-gamma[r_index]
print("seach space extended to {} with DDB".format(self.scalebounds))
def maximize_expanding_volume_L(self, gp_params):
"""
Expanding volume following L ~ MaxIter
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# consider the expansion step
# backup the previous bounds
self.bounds_bk = self.bounds.copy()
self.scalebounds_bk = self.scalebounds.copy()
# the region considered is computed as follows: NewVol~OldVol*T/t
# alternatively, we compute the radius NewL~Oldl*pow(T/t,1/d)
new_radius = self.l_radius * np.power(
self.MaxIter / len(self.Y_original), 1.0 / self.dim)
# extra proportion
extra_proportion = new_radius * 1.0 / self.l_radius
#extra_radius=(new_radius-self.l_radius)/2
if extra_proportion < 1:
extra_proportion = 1
max_bounds = self.bounds.copy()
# expand half to the lower bound and half to the upper bound
max_bounds[:, 0] = max_bounds[:, 0] - self.max_min_gap * (
extra_proportion - 1) * 0.5
max_bounds[:, 1] = max_bounds[:, 1] + self.max_min_gap * (
extra_proportion - 1) * 0.5
# make sure it is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
temp = [(max_bounds[d, :] - self.bounds_bk[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new for cropping
IsCropping = 0
if IsCropping == 1:
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d,
0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d,
1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound
self.scalebounds = self.scalebounds_bk
self.bounds = self.bounds_bk.copy()
else: # inside the old bound => recompute bound
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
else:
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
def maximize_volume_doubling(self, gp_params):
"""
Volume Doubling, double the volume (e.g., gamma=2) after every 3d evaluations
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Find unique rows of X to avoid GP from breaking
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
self.scalebounds_bk = self.scalebounds.copy()
self.bounds_bk = self.bounds
# consider the expansion step after 3 iterations
if (len(self.Y) % 3) == 0:
new_radius = 2.0 * self.l_radius
extra_radius = (new_radius - self.l_radius) / 2
max_bounds = self.bounds.copy()
max_bounds[:, 0] = max_bounds[:, 0] - extra_radius
max_bounds[:, 1] = max_bounds[:, 1] + extra_radius
# make sure it is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
self.bounds = np.asarray(max_bounds).copy()
temp = [(max_bounds[d, :] - self.bounds_bk[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
try:
self.gp.fit(self.X[ur], self.Y[ur])
except:
print "bug"
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
def maximize_unbounded_regularizer(self, gp_params):
"""
Unbounded Regularizer AISTAST 2016 Bobak
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Find unique rows of X to avoid GP from breaking
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
self.scalebounds_bk = self.scalebounds.copy()
self.bounds_bk = self.bounds
# consider the expansion step after 3 iterations
if (len(self.Y) % 3) == 0:
new_radius = 2.0 * self.l_radius
extra_radius = (new_radius - self.l_radius) / 2
max_bounds = self.bounds.copy()
max_bounds[:, 0] = max_bounds[:, 0] - extra_radius
max_bounds[:, 1] = max_bounds[:, 1] + extra_radius
# make sure it is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
self.bounds = np.asarray(max_bounds)
temp = [(max_bounds[d, :] - self.bounds[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# select the acquisition function
self.acq['x_bar'] = np.mean(self.bounds)
self.acq['R'] = np.power(self.l_radius, 1.0 / self.dim)
self.acq_func = AcquisitionFunction(self.acq)
# mean of the domain
#acq['R']
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d, 0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d, 1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound
self.scalebounds = self.scalebounds_bk
else: # inside the old bound => recompute bound
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
class PradaBayOptFBO(object):
def __init__(self,
gp_params,
f,
b_init_lower,
b_init_upper,
b_limit_lower,
b_limit_upper,
acq,
verbose=1,
opt_toolbox='nlopt'):
"""
Input parameters
----------
f: function to optimize:
pbounds0: bounds on parameters predefined
acq: acquisition function, acq['name']=['ei','ucb','poi','lei']
,acq['kappa'] for ucb, acq['k'] for lei
opt: optimization toolbox, 'nlopt','direct','scipy'
Returns
-------
dim: dimension
bounds0: initial bounds on original scale
bounds_limit: limit bounds on original scale
bounds: bounds on parameters (current)
bounds_list: bounds at all iterations
bounds_bk: bounds backup for computational purpose
scalebounds: bounds on normalized scale of 0-1 # be careful with scaling
scalebounds_bk: bounds on normalized scale of 0-1 backup for computation
time_opt: will record the time spent on optimization
gp: Gaussian Process object
MaxIter: Maximum number of iterations
"""
# Find number of parameters
self.dim = len(b_init_lower)
self.b_init_lower = b_init_lower
self.b_init_upper = b_init_upper
self.bounds0 = np.asarray([b_init_lower, b_init_upper]).T
self.bounds = self.bounds0.copy()
self.bounds_list = self.bounds0.copy()
self.bounds_bk = self.bounds.copy() # keep track
# create a scalebounds 0-1
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
self.max_min_gap_bk = self.max_min_gap.copy()
# Some function to be optimized
self.f = f
# optimization toolbox
self.opt_toolbox = opt_toolbox
# acquisition function type
self.acq = acq
# store X in original scale
self.X_original = None
# store X in 0-1 scale
self.X = None
# store y=f(x)
# (y - mean)/(max-min)
self.Y = None
# y original scale
self.Y_original = None
self.time_opt = 0
self.k_Neighbor = 2
# Lipschitz constant
self.L = 0
# Gaussian Process class
self.gp = PradaGaussianProcess(gp_params)
# acquisition function
self.acq_func = None
# stop condition
self.stop_flag = 0
# volume of initial box
# compute in log space
#self.vol0=prod(self.max_min_gap)
self.l_radius0 = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
self.l_radius = self.l_radius0
self.MaxIter = gp_params['MaxIter']
self.b_limit_lower = b_limit_lower
self.b_limit_upper = b_limit_upper
# visualization purpose
self.X_invasion = []
# will be later used for visualization
def posterior(self, Xnew):
self.gp.fit(self.X, self.Y)
mu, sigma2 = self.gp.predict(Xnew, eval_MSE=True)
return mu, np.sqrt(sigma2)
def init(self, gp_params, n_init_points=3):
"""
Input parameters
----------
gp_params: Gaussian Process structure
n_init_points: # init points
"""
# Generate random points
l = [
np.random.uniform(x[0], x[1], size=n_init_points)
for x in self.bounds
]
# Concatenate new random points to possible existing
# points from self.explore method.
temp = np.asarray(l)
temp = temp.T
init_X = list(temp.reshape((n_init_points, -1)))
self.X_original = np.asarray(init_X)
# Evaluate target function at all initialization
y_init = self.f(init_X)
y_init = np.reshape(y_init, (n_init_points, 1))
self.Y_original = np.asarray(y_init)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.X = self.X_original.copy()
def max_volume(self, gp, max_bounds_scale, max_lcb):
"""
A function to find the a data point that maximums a searching volume
Input Parameters
----------
ac: The acquisition function object that return its point-wise value.
gp: A gaussian process fitted to the relevant data.
y_max: The current maximum known value of the target function.
bounds: The variables bounds to limit the search of the acq max.
Returns
-------
x_max, The arg max of the acquisition function.
"""
def compute_utility_score_for_maximizing_volume_wrapper(
x_tries, gp, dim, max_lcb):
if len(x_tries.shape) == 1:
return compute_utility_score_for_maximizing_volume(
x_tries, gp, dim, max_lcb)
return np.apply_along_axis(
compute_utility_score_for_maximizing_volume, 1, x_tries, gp,
dim, max_lcb)
def compute_utility_score_for_maximizing_volume(
x_tries, gp, dim, max_lcb):
new_bounds = self.scalebounds
kappa = 2
mean, var = gp.predict(x_tries, eval_MSE=True)
var.flags['WRITEABLE'] = True
#var=var.copy()
var[var < 1e-10] = 0
myucb = mean + kappa * np.sqrt(var)
myucb = np.ravel(myucb)
if np.asscalar(myucb) < np.asscalar(max_lcb):
return myucb
# store the points (outside the previous bound) that sastify the constraint (original scale)
# convert to original scale before adding to
x_tries_original = x_tries * self.max_min_gap + self.bounds_bk[:,
0]
# check if it is outside the old bound
flagOutside = 0
for d in xrange(self.dim):
if x_tries[d] > self.scalebounds_bk[d, 1] or x_tries[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
break
if flagOutside == 1: # append to the invasion set
if len(self.X_invasion) == 0:
self.X_invasion = x_tries_original
self.Y_invasion = myucb
else:
self.X_invasion = np.vstack(
(self.X_invasion, x_tries_original))
self.Y_invasion = np.vstack((self.Y_invasion, myucb))
# expanse the bound
for d in xrange(dim):
# expand lower bound
if x_tries[d] < new_bounds[d, 0]:
new_bounds[d, 0] = x_tries[d]
if x_tries[d] > new_bounds[d, 1]:
new_bounds[d, 1] = x_tries[d]
self.scalebounds = new_bounds
# update the utility score
return myucb
dim = max_bounds_scale.shape[0]
# Start with the lower bound as the argmax
#x_max = max_bounds[:, 0]
max_acq = None
myopts = {'maxiter': 1000, 'fatol': 0.001, 'xatol': 0.001}
# multi start
for i in xrange(5 * dim):
# Find the minimum of minus the acquisition function
x_tries = np.random.uniform(max_bounds_scale[:, 0],
max_bounds_scale[:, 1],
size=(100 * dim, dim))
# evaluate L(x)
# estimate new L
y_tries = compute_utility_score_for_maximizing_volume_wrapper(
x_tries, gp, dim, max_lcb)
#find x optimal for init
idx_max = np.argmax(y_tries)
x_init_max = x_tries[idx_max]
res = minimize(
lambda x: -compute_utility_score_for_maximizing_volume_wrapper(
x, gp, dim, max_lcb),
#x_init_max.reshape(1, -1),bounds=bounds,options=myopts,method="nelder-mead")#L-BFGS-B
x_init_max.reshape(1, -1),
bounds=max_bounds_scale,
options=myopts,
method="L-BFGS-B") #L-BFGS-B
# value at the estimated point
val = compute_utility_score_for_maximizing_volume(
res.x, gp, dim, max_lcb)
# Store it if better than previous minimum(maximum).
if max_acq is None or val >= max_acq:
x_max = res.x
max_acq = val
#print max_acq
# Clip output to make sure it lies within the bounds. Due to floating
# point technicalities this is not always the case.
def run_FBO(self, gp_params):
"""
Main optimization method for filtering strategy for BO.
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
# for random approach
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
# obtain the maximum on the observed set (for EI)
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# consider the expansion step
# finding the maximum over the lower bound
# mu(x)-kappa x sigma(x)
mu_acq = {}
mu_acq['name'] = 'lcb'
mu_acq['dim'] = self.dim
mu_acq['kappa'] = 2
acq_mu = AcquisitionFunction(mu_acq)
# obtain the argmax(lcb), make sure the scale bound vs original bound
x_lcb_max = acq_max(ac=acq_mu.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
# obtain the max(lcb)
max_lcb = acq_mu.acq_kind(x_lcb_max, gp=self.gp, y_max=y_max)
max_lcb = np.ravel(max_lcb)
# finding the region outside the box, that has the ucb > max_lcb
self.max_min_gap_bk = self.max_min_gap.copy()
self.bounds_bk = self.bounds.copy()
self.scalebounds_bk = self.scalebounds.copy()
self.X_invasion = []
# the region considered is computed as follows: NewVol~OldVol*T/t
# alternatively, we compute the radius NewL~Oldl*pow(T/t,1/d)
new_radius = self.l_radius * np.power(
self.MaxIter / len(self.Y_original), 1.0 / self.dim)
# extra proportion
extra_proportion = new_radius * 1.0 / self.l_radius
#extra_radius=(new_radius-self.l_radius)/2
# check if extra radius is negative
if extra_proportion < 1:
extra_proportion = 1
max_bounds = self.bounds.copy()
# expand half to the lower bound and half to the upper bound, X'_t
max_bounds[:, 0] = max_bounds[:, 0] - self.max_min_gap * (
extra_proportion - 1)
max_bounds[:, 1] = max_bounds[:, 1] + self.max_min_gap * (
extra_proportion - 1)
#max_bounds[:,0]=max_bounds[:,0]-extra_radius
#max_bounds[:,1]=max_bounds[:,1]+extra_radius
# make sure the max_bounds is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
temp = [
(max_bounds[d, :] - self.bounds[d, 0]) * 1.0 / self.max_min_gap[d]
for d in xrange(self.dim)
]
max_bounds_scale = np.asarray(temp)
# find suitable candidates in new regions
# ucb(x) > max_lcb st max L(x)
# new bound in scale space
# we note that the scalebound will be changed inside this function
self.max_volume(self.gp, max_bounds_scale, max_lcb)
#print "new bounds scale"
#print self.scalebounds
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
val_acq = self.acq_func.acq_kind(x_max_scale, self.gp, y_max)
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d, 0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d, 1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
else:
self.scalebounds[d, :] = self.scalebounds_bk[d, :]
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound, use the old bound
self.scalebounds = self.scalebounds_bk
self.bounds = self.bounds_bk.copy()
else: # outside the old bound => expand the bound as the minimum bound containing the old bound and the selected point
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
self.bounds_list = np.hstack((self.bounds_list, self.bounds))
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
#======================================================================================
#======================================================================================================
#======================================================================================================
#======================================================================================================
def maximize_volume_doubling(self, gp_params):
"""
Volume Doubling, double the volume (e.g., gamma=2) after every 3d evaluations
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Find unique rows of X to avoid GP from breaking
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
self.scalebounds_bk = self.scalebounds.copy()
self.bounds_bk = self.bounds
# consider the expansion step after 3 iterations
if (len(self.Y) % 3) == 0:
new_radius = 2.0 * self.l_radius
extra_radius = (new_radius - self.l_radius) / 2
max_bounds = self.bounds.copy()
max_bounds[:, 0] = max_bounds[:, 0] - extra_radius
max_bounds[:, 1] = max_bounds[:, 1] + extra_radius
# make sure it is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
self.bounds = np.asarray(max_bounds).copy()
temp = [(max_bounds[d, :] - self.bounds_bk[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
try:
self.gp.fit(self.X[ur], self.Y[ur])
except:
print "bug"
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
def maximize_unbounded_regularizer(self, gp_params):
"""
Unbounded Regularizer AISTAST 2016 Bobak
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Find unique rows of X to avoid GP from breaking
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
self.scalebounds_bk = self.scalebounds.copy()
self.bounds_bk = self.bounds
# consider the expansion step after 3 iterations
if (len(self.Y) % 3) == 0:
new_radius = 2.0 * self.l_radius
extra_radius = (new_radius - self.l_radius) / 2
max_bounds = self.bounds.copy()
max_bounds[:, 0] = max_bounds[:, 0] - extra_radius
max_bounds[:, 1] = max_bounds[:, 1] + extra_radius
# make sure it is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
self.bounds = np.asarray(max_bounds)
temp = [(max_bounds[d, :] - self.bounds[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# select the acquisition function
self.acq['x_bar'] = np.mean(self.bounds)
self.acq['R'] = np.power(self.l_radius, 1.0 / self.dim)
self.acq_func = AcquisitionFunction(self.acq)
# mean of the domain
#acq['R']
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d, 0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d, 1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound
self.scalebounds = self.scalebounds_bk
else: # inside the old bound => recompute bound
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
def maximize_expanding_volume_L(self, gp_params):
"""
Expanding volume following L ~ MaxIter
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# consider the expansion step
# backup the previous bounds
self.bounds_bk = self.bounds.copy()
self.scalebounds_bk = self.scalebounds.copy()
# the region considered is computed as follows: NewVol~OldVol*T/t
# alternatively, we compute the radius NewL~Oldl*pow(T/t,1/d)
new_radius = self.l_radius * np.power(
self.MaxIter / len(self.Y_original), 1.0 / self.dim)
# extra proportion
extra_proportion = new_radius * 1.0 / self.l_radius
#extra_radius=(new_radius-self.l_radius)/2
if extra_proportion < 1:
extra_proportion = 1
max_bounds = self.bounds.copy()
# expand half to the lower bound and half to the upper bound
max_bounds[:, 0] = max_bounds[:, 0] - self.max_min_gap * (
extra_proportion - 1) * 0.5
max_bounds[:, 1] = max_bounds[:, 1] + self.max_min_gap * (
extra_proportion - 1) * 0.5
# make sure it is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
temp = [(max_bounds[d, :] - self.bounds_bk[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new for cropping
IsCropping = 0
if IsCropping == 1:
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d,
0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d,
1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound
self.scalebounds = self.scalebounds_bk
self.bounds = self.bounds_bk.copy()
else: # inside the old bound => recompute bound
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
else:
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
def maximize_expanding_volume_L_Cropping(self, gp_params):
"""
Expanding volume following L ~ MaxIter
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# consider the expansion step
# finding the region outside the box, that has the ucb > max_lcb
self.bounds_bk = self.bounds.copy()
self.scalebounds_bk = self.scalebounds.copy()
# the region considered is computed as follows: NewVol~OldVol*T/t
# alternatively, we compute the radius NewL~Oldl*pow(T/t,1/d)
new_radius = self.l_radius * np.power(
self.MaxIter / len(self.Y_original), 1.0 / self.dim)
# extra proportion
extra_proportion = new_radius * 1.0 / self.l_radius
#extra_radius=(new_radius-self.l_radius)/2
# check if extra radius is negative
#if extra_radius<0:
#extra_radius=0
max_bounds = self.bounds.copy()
# expand half to the lower bound and half to the upper bound
max_bounds[:,
0] = max_bounds[:, 0] - self.max_min_gap * extra_proportion
max_bounds[:,
1] = max_bounds[:, 1] + self.max_min_gap * extra_proportion
# make sure the max_bound is still within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
temp = [(max_bounds[d, :] - self.bounds_bk[d, 0]) * 1.0 /
self.max_min_gap[d] for d in xrange(self.dim)]
self.scalebounds = np.asarray(temp)
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
#val_acq=self.acq_func.acq_kind(x_max_scale,self.gp,y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new for cropping
IsCropping = 1
if IsCropping == 1:
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d,
0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d,
1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
else:
self.scalebounds[d, :] = self.scalebounds_bk[d, :]
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound
self.scalebounds = self.scalebounds_bk
self.bounds = self.bounds_bk.copy()
else: # inside the old bound => recompute bound
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
else:
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
def run_FBO(self, gp_params):
"""
Main optimization method for filtering strategy for BO.
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
# for random approach
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# scale the data before updating the GP
# convert it to scaleX
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
# obtain the maximum on the observed set (for EI)
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# consider the expansion step
# finding the maximum over the lower bound
# mu(x)-kappa x sigma(x)
mu_acq = {}
mu_acq['name'] = 'lcb'
mu_acq['dim'] = self.dim
mu_acq['kappa'] = 2
acq_mu = AcquisitionFunction(mu_acq)
# obtain the argmax(lcb), make sure the scale bound vs original bound
x_lcb_max = acq_max(ac=acq_mu.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
# obtain the max(lcb)
max_lcb = acq_mu.acq_kind(x_lcb_max, gp=self.gp, y_max=y_max)
max_lcb = np.ravel(max_lcb)
# finding the region outside the box, that has the ucb > max_lcb
self.max_min_gap_bk = self.max_min_gap.copy()
self.bounds_bk = self.bounds.copy()
self.scalebounds_bk = self.scalebounds.copy()
self.X_invasion = []
# the region considered is computed as follows: NewVol~OldVol*T/t
# alternatively, we compute the radius NewL~Oldl*pow(T/t,1/d)
new_radius = self.l_radius * np.power(
self.MaxIter / len(self.Y_original), 1.0 / self.dim)
# extra proportion
extra_proportion = new_radius * 1.0 / self.l_radius
#extra_radius=(new_radius-self.l_radius)/2
# check if extra radius is negative
if extra_proportion < 1:
extra_proportion = 1
max_bounds = self.bounds.copy()
# expand half to the lower bound and half to the upper bound, X'_t
max_bounds[:, 0] = max_bounds[:, 0] - self.max_min_gap * (
extra_proportion - 1)
max_bounds[:, 1] = max_bounds[:, 1] + self.max_min_gap * (
extra_proportion - 1)
#max_bounds[:,0]=max_bounds[:,0]-extra_radius
#max_bounds[:,1]=max_bounds[:,1]+extra_radius
# make sure the max_bounds is within the limit
if not (self.b_limit_lower is None):
temp_max_bounds_lower = [
np.maximum(max_bounds[idx, 0], self.b_limit_lower[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 0] = temp_max_bounds_lower
if not (self.b_limit_upper is None):
temp_max_bounds_upper = [
np.minimum(max_bounds[idx, 1], self.b_limit_upper[idx])
for idx in xrange(self.dim)
]
max_bounds[:, 1] = temp_max_bounds_upper
temp = [
(max_bounds[d, :] - self.bounds[d, 0]) * 1.0 / self.max_min_gap[d]
for d in xrange(self.dim)
]
max_bounds_scale = np.asarray(temp)
# find suitable candidates in new regions
# ucb(x) > max_lcb st max L(x)
# new bound in scale space
# we note that the scalebound will be changed inside this function
self.max_volume(self.gp, max_bounds_scale, max_lcb)
#print "new bounds scale"
#print self.scalebounds
# perform standard BO on the new bound (scaled)
x_max_scale = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
val_acq = self.acq_func.acq_kind(x_max_scale, self.gp, y_max)
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max_scale).sum(axis=1) == 0):
x_max_scale = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# check if the estimated data point is in the old bound or new
flagOutside = 0
for d in xrange(self.dim):
if x_max_scale[d] > self.scalebounds_bk[d, 1] or x_max_scale[
d] < self.scalebounds_bk[d, 0]: #outside the old bound
flagOutside = 1
self.scalebounds[d, 0] = np.minimum(x_max_scale[d],
self.scalebounds_bk[d, 0])
self.scalebounds[d, 1] = np.maximum(x_max_scale[d],
self.scalebounds_bk[d, 1])
else:
self.scalebounds[d, :] = self.scalebounds_bk[d, :]
# now the scalebounds is no longer 0-1
if flagOutside == 0: # not outside the old bound, use the old bound
self.scalebounds = self.scalebounds_bk
self.bounds = self.bounds_bk.copy()
else: # outside the old bound => expand the bound as the minimum bound containing the old bound and the selected point
temp = [
self.scalebounds[d, :] * self.max_min_gap[d] +
self.bounds_bk[d, 0] for d in xrange(self.dim)
]
if self.dim > 1:
self.bounds = np.reshape(temp, (self.dim, 2))
else:
self.bounds = np.array(temp)
self.bounds_list = np.hstack((self.bounds_list, self.bounds))
# compute X in original scale
temp_X_new_original = x_max_scale * self.max_min_gap + self.bounds_bk[:,
0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# clone the self.X for updating GP
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
temp = np.divide((self.X_original - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
# evaluate Y using original X
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# for plotting
self.gp = PradaGaussianProcess(gp_params)
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# update volume and radius
#self.vol=prod(self.max_min_gap)
#self.l_radius=np.power(self.vol,1/self.dim)
self.l_radius = np.exp(1.0 * np.sum(np.log(self.max_min_gap)) /
self.dim)
class PradaBayOptBatch(object):
def __init__(self,
gp_params,
f,
pbounds,
acq,
verbose=1,
opt_toolbox='scipy'):
"""
Input parameters
----------
f: function to optimize:
pbounds: bounds on parameters
acq: acquisition function, 'ei', 'ucb'
opt: optimization toolbox, 'nlopt','direct','scipy'
Returns
-------
dim: dimension
bounds: bounds on original scale
scalebounds: bounds on normalized scale of 0-1
time_opt: will record the time spent on optimization
gp: Gaussian Process object
"""
# Store the original dictionary
self.pbounds = pbounds
# Find number of parameters
self.dim = len(pbounds)
# Create an array with parameters bounds
if isinstance(pbounds, dict):
# Get the name of the parameters
self.keys = list(pbounds.keys())
self.bounds = []
for key in self.pbounds.keys():
self.bounds.append(self.pbounds[key])
self.bounds = np.asarray(self.bounds)
else:
self.bounds = np.asarray(pbounds)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
# Some function to be optimized
self.f = f
# optimization tool: direct, scipy, nlopt
self.opt_toolbox = opt_toolbox
# acquisition function type
self.acq = acq
# store the batch size for each iteration
self.NumPoints = []
# Numpy array place holders
self.X_original = None
# scale the data to 0-1 fit GP better
self.X = None # X=( X_original - min(bounds) / (max(bounds) - min(bounds))
self.Y = None # Y=( Y_original - mean(bounds) / (max(bounds) - min(bounds))
self.Y_original = None
self.opt_time = 0
self.L = 0 # lipschitz
self.gp = PradaGaussianProcess(gp_params)
# Acquisition Function
#self.acq_func = None
self.acq_func = AcquisitionFunction(acq=self.acq)
def posterior(self, Xnew):
#xmin, xmax = -2, 10
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
mu, sigma2 = self.gp.predict(Xnew, eval_MSE=True)
return mu, np.sqrt(sigma2)
def init(self, n_init_points):
"""
Input parameters
----------
gp_params: Gaussian Process structure
n_init_points: # init points
"""
# Generate random points
l = [
np.random.uniform(x[0], x[1], size=n_init_points)
for x in self.bounds
]
# Concatenate new random points to possible existing
# points from self.explore method.
#self.init_points += list(map(list, zip(*l)))
temp = np.asarray(l)
temp = temp.T
init_X = list(temp.reshape((n_init_points, -1)))
# Evaluate target function at all initialization
y_init = self.f(init_X)
# Turn it into np array and store.
self.X_original = np.asarray(init_X)
temp_init_point = np.divide((init_X - self.bounds[:, 0]),
self.max_min_gap)
self.X_original = np.asarray(init_X)
self.X = np.asarray(temp_init_point)
y_init = np.reshape(y_init, (n_init_points, 1))
self.Y_original = np.asarray(y_init)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.NumPoints = np.append(self.NumPoints, n_init_points)
# Set parameters if any was passed
#self.gp=PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
#ur = unique_rows(self.X)
#self.gp.fit(self.X[ur], self.Y[ur])
#print "#Batch={:d} f_max={:.4f}".format(n_init_points,self.Y.max())
def init_with_data(self, init_X, init_Y):
"""
Input parameters
----------
gp_params: Gaussian Process structure
x,y: # init data observations (in original scale)
"""
# Turn it into np array and store.
self.X_original = np.asarray(init_X)
temp_init_point = np.divide((init_X - self.bounds[:, 0]),
self.max_min_gap)
self.X_original = np.asarray(init_X)
self.X = np.asarray(temp_init_point)
self.Y_original = np.asarray(init_Y)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.NumPoints = np.append(self.NumPoints, len(init_Y))
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
def smooth_the_peak(self, my_peak):
# define the local bound around the estimated point
local_bound = np.zeros((self.dim, 2))
for dd in range(self.dim):
try:
local_bound[dd, 0] = my_peak[-1][dd] - 0.005
local_bound[dd, 1] = my_peak[-1][dd] + 0.005
except:
local_bound[dd, 0] = my_peak[dd] - 0.005
local_bound[dd, 1] = my_peak[dd] + 0.005
local_bound = np.clip(local_bound, self.scalebounds[:, 0],
self.scalebounds[:, 1])
dim = len(local_bound)
num_data = 1000 * dim
samples = np.zeros(shape=(num_data, dim))
#for k in range(0,dim): samples[:,k] = np.random.uniform(low=local_bound[k][0],high=local_bound[k][1],size=num_data)
for dd in range(0, dim):
samples[:, dd] = np.linspace(local_bound[dd][0],
local_bound[dd][1], num_data)
# smooth the peak
"""
n_bins = 100*np.ones(self.dim)
mygrid = np.mgrid[[slice(row[0], row[1], n*1j) for row, n in zip(local_bound, n_bins)]]
mygrid=mygrid.reshape(100**self.dim, self.dim)
utility_grid=self.acq_func.acq_kind(mygrid,self.gp,self.Y.max())
mysamples=np.vstack((mygrid,utility_grid))
samples_smooth=filters.uniform_filter(mysamples, size=[2,2], output=None, mode='reflect', cval=0.0, origin=0)
"""
# get the utility after smoothing
samples_smooth = samples
utility_smooth = self.acq_func.acq_kind(samples_smooth, self.gp,
self.Y.max())
# get the peak value y
#peak_y=np.max(utility_smooth)
# get the peak location x
#peak_x=samples_smooth[np.argmax(utility_smooth)]
peak_x = my_peak
# linear regression
regr = linear_model.LinearRegression()
regr.fit(samples_smooth, utility_smooth)
#residual_ss=np.mean((regr.predict(samples_smooth) - utility_smooth) ** 2)
mystd = np.std(utility_smooth)
return peak_x, mystd
def check_real_peak(self, my_peak, threshold=0.1):
# define the local bound around the estimated point
local_bound = np.zeros((self.dim, 2))
for dd in range(self.dim):
try:
local_bound[dd, 0] = my_peak[-1][dd] - 0.01
local_bound[dd, 1] = my_peak[-1][dd] + 0.01
except:
local_bound[dd, 0] = my_peak[dd] - 0.01
local_bound[dd, 1] = my_peak[dd] + 0.01
#local_bound=np.clip(local_bound,self.scalebounds[:,0],self.scalebounds[:,1])
local_bound[:, 0] = local_bound[:, 0].clip(self.scalebounds[:, 0],
self.scalebounds[:, 1])
local_bound[:, 1] = local_bound[:, 1].clip(self.scalebounds[:, 0],
self.scalebounds[:, 1])
dim = len(local_bound)
num_data = 100 * dim
samples = np.zeros(shape=(num_data, dim))
for dd in range(0, dim):
samples[:, dd] = np.linspace(local_bound[dd][0],
local_bound[dd][1], num_data)
# get the utility after smoothing
myutility = self.acq_func.acq_kind(samples, self.gp, self.Y.max())
# linear regression
#regr = linear_model.LinearRegression()
#regr.fit(samples, myutility)
#residual_ss=np.mean((regr.predict(samples_smooth) - utility_smooth) ** 2)
#mystd=np.std(myutility)
mystd = np.mean(myutility)
IsPeak = 0
if mystd > threshold / (self.dim**2):
IsPeak = 1
return IsPeak, mystd
def estimate_L(self, bounds):
'''
Estimate the Lipschitz constant of f by taking maximizing the norm of the expectation of the gradient of *f*.
'''
def df(x, model, x0):
mean_derivative = gp_model.predictive_gradient(self.X, self.Y, x)
temp = mean_derivative * mean_derivative
if len(temp.shape) <= 1:
res = np.sqrt(temp)
else:
res = np.sqrt(
np.sum(temp, axis=1)
) # simply take the norm of the expectation of the gradient
return -res
gp_model = self.gp
dim = len(bounds)
num_data = 1000 * dim
samples = np.zeros(shape=(num_data, dim))
for k in range(0, dim):
samples[:, k] = np.random.uniform(low=bounds[k][0],
high=bounds[k][1],
size=num_data)
#samples = np.vstack([samples,gp_model.X])
pred_samples = df(samples, gp_model, 0)
x0 = samples[np.argmin(pred_samples)]
res = minimize(df,
x0,
method='L-BFGS-B',
bounds=bounds,
args=(gp_model, x0),
options={'maxiter': 100})
try:
minusL = res.fun[0][0]
except:
if len(res.fun.shape) == 1:
minusL = res.fun[0]
else:
minusL = res.fun
L = -minusL
if L < 1e-6:
L = 0.0001 ## to avoid problems in cases in which the model is flat.
return L
def maximize_batch_PS(self, gp_params, B=5, kappa=2):
"""
Finding a batch of points using Peak Suppression approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
kappa: constant value in UCB
Returns
-------
X: a batch of [x_1..x_Nt]
"""
const_liar = self.Y_original.min()
# Set acquisition function
#self.acq_func = AcquisitionFunction(kind=self.acq, kappa=kappa)
y_max = self.Y.max()
# Set parameters if any was passed
self.gp = PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
start_opt = time.time()
# copy GP, X and Y
temp_gp = self.gp
temp_X = self.X
temp_Y = self.Y
#store new_x
new_X = []
stdPeak = [0] * B
IsPeak = [0] * B
for ii in range(B):
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=temp_gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
# Test if x_max is repeated, if it is, draw another one at random
if np.any((np.abs(temp_X - x_max)).sum(axis=1) < 0.002 *
self.dim) | np.isnan(x_max.sum()):
#x_max = np.random.uniform(self.scalebounds[:, 0], self.scalebounds[:, 1],size=self.scalebounds.shape[0])
IsPeak[ii] = 0
stdPeak[ii] = 0
print "reject"
else:
IsPeak[ii], stdPeak[ii] = self.check_real_peak(x_max)
print "IsPeak={:d} std={:.5f}".format(IsPeak[ii], stdPeak[ii])
if ii == 0:
new_X = x_max
else:
new_X = np.vstack((new_X, x_max.reshape((1, -1))))
temp_X = np.vstack((temp_X, x_max.reshape((1, -1))))
temp_Y = np.append(temp_Y, const_liar)
#temp_gp.fit(temp_X,temp_Y)
temp_gp.fit_incremental(x_max, np.asarray([const_liar]))
"""
toplot_bo=copy.deepcopy(self)
toplot_bo.gp=copy.deepcopy(temp_gp)
toplot_bo.X=temp_X
toplot_bo.X_original=[val*self.max_min_gap+self.bounds[:,0] for idx, val in enumerate(temp_X)]
toplot_bo.X_original=np.asarray(toplot_bo.X_original)
toplot_bo.Y=temp_Y
toplot_bo.Y_original=temp_Y*(np.max(self.Y_original)-np.min(self.Y_original))+np.mean(self.Y_original)
visualization.plot_bo(toplot_bo)
"""
IsPeak = np.asarray(IsPeak)
# check if there is no real peak, then pick up the top peak (highest std)
# rank the peak
idx = np.sort(stdPeak)
if np.sum(IsPeak) == 0:
top_peak = np.argmax(stdPeak)
new_X = new_X[top_peak]
else:
new_X = new_X[IsPeak == 1]
print new_X
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.opt_time = np.hstack((self.opt_time, elapse_opt))
# Updating the GP.
#new_X=new_X.reshape((-1, self.dim))
# Test if x_max is repeated, if it is, draw another one at random
temp_new_X = []
for idx, val in enumerate(new_X):
if np.all(
np.any(np.abs(self.X - val) > 0.02,
axis=1)): # check if a data point is already taken
temp_new_X = np.append(temp_new_X, val)
if len(temp_new_X) == 0:
temp_new_X = np.zeros((1, self.dim))
for idx in range(0, self.dim):
temp_new_X[0,
idx] = np.random.uniform(self.scalebounds[idx, 0],
self.scalebounds[idx,
1], 1)
else:
temp_new_X = temp_new_X.reshape((-1, self.dim))
self.X = np.vstack((self.X, temp_new_X))
# convert back to original scale
temp_X_new_original = [
val * self.max_min_gap + self.bounds[:, 0]
for idx, val in enumerate(temp_new_X)
]
temp_X_new_original = np.asarray(temp_X_new_original)
self.X_original = np.vstack((self.X_original, temp_X_new_original))
for idx, val in enumerate(temp_X_new_original):
self.Y_original = np.append(self.Y_original, self.f(val))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.NumPoints = np.append(self.NumPoints,
temp_X_new_original.shape[0])
print "#Batch={:d} f_max={:.4f}".format(temp_X_new_original.shape[0],
self.Y_original.max())
def maximize_batch_CL(self, gp_params, B=5):
"""
Finding a batch of points using Constant Liar approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
kappa: constant value in UCB
Returns
-------
X: a batch of [x_1..x_Nt]
"""
self.NumPoints = np.append(self.NumPoints, B)
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=B) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.opt_time = np.hstack((self.opt_time, 0))
return
#const_liar=self.Y.mean()
#const_liar=self.Y_original.mean()
#const_liar=self.Y.max()
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
y_max = self.Y.max()
# Set parameters if any was passed
self.gp = PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
start_opt = time.time()
# copy GP, X and Y
temp_gp = self.gp
temp_X = self.X
temp_Y = self.Y
#temp_Y_original=self.Y_original
#store new_x
new_X = []
for ii in range(B):
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=temp_gp,
y_max=y_max,
bounds=self.scalebounds)
val_acq = self.acq_func.acq_kind(x_max, temp_gp, y_max)
print "CL alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# Test if x_max is repeated, if it is, draw another one at random
# If it is repeated, print a warning
#if np.any((self.X - x_max).sum(axis=1) == 0) | np.isnan(x_max.sum()):
#x_max = np.random.uniform(self.scalebounds[:, 0], self.scalebounds[:, 1],size=self.scalebounds.shape[0])
if ii == 0:
new_X = x_max
else:
new_X = np.vstack((new_X, x_max.reshape((1, -1))))
temp_X = np.vstack((temp_X, x_max.reshape((1, -1))))
const_liar, const_liar_variance = temp_gp.predict(x_max,
eval_MSE=1)
temp_Y = np.append(temp_Y, const_liar)
temp_gp.fit(temp_X, temp_Y)
# Updating the GP.
new_X = new_X.reshape((B, -1))
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.opt_time = np.hstack((self.opt_time, elapse_opt))
#print new_X
self.X = np.vstack((self.X, new_X))
# convert back to original scale
temp_X_new_original = [
val * self.max_min_gap + self.bounds[:, 0]
for idx, val in enumerate(new_X)
]
temp_X_new_original = np.asarray(temp_X_new_original)
self.X_original = np.vstack((self.X_original, temp_X_new_original))
for idx, val in enumerate(temp_X_new_original):
self.Y_original = np.append(self.Y_original, self.f(val))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
#print "#Batch={:d} f_max={:.4f}".format(B,self.Y_original.max())
return new_X, temp_X_new_original
def maximize_batch_CL_incremental(self, gp_params, B=5):
"""
Finding a batch of points using Constant Liar approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
kappa: constant value in UCB
Returns
-------
X: a batch of [x_1..x_Nt]
"""
self.NumPoints = np.append(self.NumPoints, B)
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=B) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.opt_time = np.hstack((self.opt_time, 0))
return
#const_liar=self.Y.mean()
#const_liar=self.Y_original.min()
#const_liar=self.Y.max()
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
y_max = self.Y.max()
# Set parameters if any was passed
self.gp = PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
start_opt = time.time()
# copy GP, X and Y
temp_gp = copy.deepcopy(self.gp)
temp_X = self.X
temp_Y = self.Y
#temp_Y_original=self.Y_original
#store new_x
new_X = []
for ii in range(B):
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=temp_gp,
y_max=y_max,
bounds=self.scalebounds)
# Test if x_max is repeated, if it is, draw another one at random
if np.any(
np.any(np.abs(self.X - x_max) < 0.02,
axis=1)): # check if a data point is already taken
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
if ii == 0:
new_X = x_max
else:
new_X = np.vstack((new_X, x_max.reshape((1, -1))))
const_liar = temp_gp.predict(x_max, eval_MSE=true)
#temp_X= np.vstack((temp_X, x_max.reshape((1, -1))))
#temp_Y = np.append(temp_Y, const_liar )
#temp_gp.fit(temp_X,temp_Y)
# update the Gaussian Process and thus the acquisition function
#temp_gp.compute_incremental_var(temp_X,x_max)
temp_gp.fit_incremental(x_max, np.asarray([const_liar]))
# Updating the GP.
new_X = new_X.reshape((B, -1))
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.opt_time = np.hstack((self.opt_time, elapse_opt))
#print new_X
self.X = np.vstack((self.X, new_X))
# convert back to original scale
temp_X_new_original = [
val * self.max_min_gap + self.bounds[:, 0]
for idx, val in enumerate(new_X)
]
temp_X_new_original = np.asarray(temp_X_new_original)
self.X_original = np.vstack((self.X_original, temp_X_new_original))
for idx, val in enumerate(temp_X_new_original):
self.Y_original = np.append(self.Y_original, self.f(val))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
#print "#Batch={:d} f_max={:.4f}".format(B,self.Y_original.max())
def fitIGMM(self, obs, IsPlot=0):
"""
Fitting the Infinite Gaussian Mixture Model and GMM where applicable
Input Parameters
----------
obs: samples generated under the acqusition function by BGSS
IsPlot: flag variable for visualization
Returns
-------
mean vector: mu_1,...mu_K
"""
if self.dim <= 2:
n_init_components = 3
else:
n_init_components = np.int(self.dim * 1.1)
dpgmm = mixture.DPGMM(n_components=n_init_components,
covariance_type="full",
min_covar=10)
dpgmm.fit(obs)
# check if DPGMM fail, then use GMM.
mydist = euclidean_distances(dpgmm.means_, dpgmm.means_)
np.fill_diagonal(mydist, 99)
if dpgmm.converged_ is False or np.min(mydist) < (0.01 * self.dim):
dpgmm = mixture.GMM(n_components=n_init_components,
covariance_type="full",
min_covar=1e-3)
dpgmm.fit(obs)
if self.dim >= 5:
# since kmeans does not provide weight and means, we will manually compute it
try:
dpgmm.weights_ = np.histogram(dpgmm.labels_,
np.int(self.dim * 1.2))
dpgmm.weights_ = np.true_divide(dpgmm.weights_[0],
np.sum(dpgmm.weights_[0]))
dpgmm.means_ = dpgmm.cluster_centers_
except:
pass
# truncated for variational inference
weight = dpgmm.weights_
weight_sorted = np.sort(weight)
weight_sorted = weight_sorted[::-1]
temp_cumsum = np.cumsum(weight_sorted)
cutpoint = 0
for idx, val in enumerate(temp_cumsum):
if val > 0.73:
cutpoint = weight_sorted[idx]
break
ClusterIndex = [
idx for idx, val in enumerate(dpgmm.weights_) if val >= cutpoint
]
myMeans = dpgmm.means_[ClusterIndex]
#dpgmm.means_=dpgmm.means_[ClusterIndex]
dpgmm.truncated_means_ = dpgmm.means_[ClusterIndex]
#myCov=dpgmm.covars_[ClusterIndex]
if IsPlot == 1 and self.dim <= 2:
visualization.plot_histogram(self, obs)
visualization.plot_mixturemodel(dpgmm, self, obs)
new_X = myMeans.reshape((len(ClusterIndex), -1))
new_X = new_X.tolist()
return new_X
def maximize_batch_B3O(self, gp_params, kappa=2, IsPlot=0):
"""
Finding a batch of points using Budgeted Batch Bayesian Optimization approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
kappa: constant value in UCB
IsPlot: flag variable for visualization
Returns
-------
X: a batch of [x_1..x_Nt]
"""
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# Step 2 in the Algorithm
# Set parameters for Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
if len(self.gp.KK_x_x_inv) == 0: # check if empty
self.gp.fit(self.X, self.Y)
#else:
#self.gp.fit_incremental(self.X[ur], self.Y[ur])
# record optimization time
start_gmm_opt = time.time()
if IsPlot == 1 and self.dim <= 2: #plot
visualization.plot_bo(self)
# Step 4 in the Algorithm
# generate samples from Acquisition function
# check the bound 0-1 or original bound
obs = acq_batch_generalized_slice_sampling_generate(
self.acq_func.acq_kind,
self.gp,
self.scalebounds,
N=500,
y_max=self.Y.max())
# Step 5 and 6 in the Algorithm
if len(obs) == 0: # monotonous acquisition function
print "Monotonous acquisition function"
new_X = np.random.uniform(self.bounds[:, 0],
self.bounds[:, 1],
size=self.bounds.shape[0])
new_X = new_X.reshape((1, -1))
new_X = new_X.tolist()
else:
new_X = self.fitIGMM(obs, IsPlot)
# Test if x_max is repeated, if it is, draw another one at random
temp_new_X = []
for idx, val in enumerate(new_X):
if np.all(
np.any(np.abs(self.X - val) > 0.02,
axis=1)): # check if a data point is already taken
temp_new_X = np.append(temp_new_X, val)
if len(temp_new_X) == 0:
temp_new_X = np.zeros((1, self.dim))
for idx in range(0, self.dim):
temp_new_X[0,
idx] = np.random.uniform(self.scalebounds[idx, 0],
self.scalebounds[idx,
1], 1)
else:
temp_new_X = temp_new_X.reshape((-1, self.dim))
self.NumPoints = np.append(self.NumPoints, temp_new_X.shape[0])
finished_gmm_opt = time.time()
elapse_gmm_opt = finished_gmm_opt - start_gmm_opt
self.opt_time = np.hstack((self.opt_time, elapse_gmm_opt))
self.X = np.vstack((self.X, temp_new_X))
temp_X_new_original = [
val * self.max_min_gap + self.bounds[:, 0]
for idx, val in enumerate(temp_new_X)
]
temp_X_new_original = np.asarray(temp_X_new_original)
# Step 7 in the algorithm
# Evaluate y=f(x)
temp = self.f(temp_X_new_original)
temp = np.reshape(temp, (-1, 1))
# Step 8 in the algorithm
self.Y_original = np.append(self.Y_original, temp)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.X_original = np.vstack((self.X_original, temp_X_new_original))
print "#Batch={:d} f_max={:.4f}".format(temp_new_X.shape[0],
self.Y_original.max())
#ur = unique_rows(self.X)
#self.gp.fit(self.X[ur], self.Y[ur])
#self.gp.fit_incremental(temp_new_X, temp_new_Y)
#======================================================================================
#======================================================================================================
#======================================================================================================
#======================================================================================================
def maximize_batch_BUCB(self, gp_params, B=5):
"""
Finding a batch of points using GP-BUCB approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
B: fixed batch size for all iteration
kappa: constant value in UCB
IsPlot: flag variable for visualization
Returns
-------
X: a batch of [x_1..x_B]
"""
self.B = B
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
# Set parameters if any was passed
self.gp = PradaGaussianProcess(gp_params)
if len(self.gp.KK_x_x_inv) == 0: # check if empty
self.gp.fit(self.X, self.Y)
#else:
#self.gp.fit_incremental(self.X[ur], self.Y[ur])
start_gmm_opt = time.time()
y_max = self.gp.Y.max()
# check the bound 0-1 or original bound
temp_X = self.X
temp_gp = self.gp
temp_gp.X_bucb = temp_X
temp_gp.KK_x_x_inv_bucb = self.gp.KK_x_x_inv
# finding new X
new_X = []
for ii in range(B):
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=temp_gp,
y_max=y_max,
bounds=self.scalebounds)
if np.any(
(temp_X - x_max).sum(axis=1) == 0) | np.isnan(x_max.sum()):
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
if ii == 0:
new_X = x_max
else:
new_X = np.vstack((new_X, x_max.reshape((1, -1))))
# update the Gaussian Process and thus the acquisition function
temp_gp.compute_incremental_var(temp_X, x_max)
temp_X = np.vstack((temp_X, x_max.reshape((1, -1))))
temp_gp.X_bucb = temp_X
# record the optimization time
finished_gmm_opt = time.time()
elapse_gmm_opt = finished_gmm_opt - start_gmm_opt
self.time_gmm_opt = np.hstack((self.time_gmm_opt, elapse_gmm_opt))
self.NumPoints = np.append(self.NumPoints, B)
self.X = temp_X
# convert back to original scale
temp_X_new_original = [
val * self.max_min_gap + self.bounds[:, 0]
for idx, val in enumerate(new_X)
]
temp_X_new_original = np.asarray(temp_X_new_original)
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# evaluate y=f(x)
temp = self.f(temp_X_new_original)
temp = np.reshape(temp, (-1, 1))
self.Y_original = np.append(self.Y_original, temp)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
print "#Batch={:d} f_max={:.4f}".format(new_X.shape[0],
self.Y_original.max())
def maximize(self,gp_params,kappa=2):
"""
Main optimization method.
Input parameters
----------
kappa: parameter for UCB acquisition only.
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
# init a new Gaussian Process
self.gp=PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt=time.time()
acq=self.acq
# select the acquisition function
if acq=='nei':
self.L=self.estimate_L(self.bounds)
self.util = AcquisitionFunction(kind=self.acq, L=self.L)
else:
if acq=="ucb":
self.acq_func = AcquisitionFunction(kind=self.acq, kappa=kappa)
else:
self.acq_func = AcquisitionFunction(kind=self.acq, kappa=kappa)
y_max = self.Y.max()
# select the optimization toolbox
if self.opt=='nlopt':
x_max,f_max = acq_max_nlopt(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
if self.opt=='scipy':
x_max = acq_max(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
if self.opt=='direct':
x_max = acq_max_direct(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
# record the optimization time
finished_opt=time.time()
elapse_opt=finished_opt-start_opt
self.time_opt=np.hstack((self.time_opt,elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max).sum(axis=1) == 0):
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# store X
self.X = np.vstack((self.X, x_max.reshape((1, -1))))
# compute X in original scale
temp_X_new_original=x_max*self.max_min_gap+self.bounds[:,0]
self.X_original=np.vstack((self.X_original, temp_X_new_original))
# evaluate Y using original X
self.Y = np.append(self.Y, self.f(temp_X_new_original))
def maximize_batch_B3O(self,gp_params, kappa=2,IsPlot=0):
"""
Finding a batch of points using Budgeted Batch Bayesian Optimization approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
kappa: constant value in UCB
IsPlot: flag variable for visualization
Returns
-------
X: a batch of [x_1..x_Nt]
"""
# Set acquisition function
self.acq_func = AcquisitionFunction(kind=self.acq, kappa=kappa)
# Step 2 in the Algorithm
# Set parameters for Gaussian Process
self.gp=PradaGaussianProcess(gp_params)
if len(self.gp.KK_x_x_inv)==0: # check if empty
self.gp.fit(self.X, self.Y)
#else:
#self.gp.fit_incremental(self.X[ur], self.Y[ur])
# record optimization time
start_gmm_opt=time.time()
if IsPlot==1 and self.dim<=2:#plot
visualization.plot_bo(self)
# Step 4 in the Algorithm
# generate samples from Acquisition function
# check the bound 0-1 or original bound
obs=acq_batch_generalized_slice_sampling_generate(self.acq_func.acq_kind,self.gp,self.scalebounds,N=500,y_max=self.Y.max())
# Step 5 and 6 in the Algorithm
if len(obs)==0: # monotonous acquisition function
print "Monotonous acquisition function"
new_X=np.random.uniform(self.bounds[:, 0],self.bounds[:, 1],size=self.bounds.shape[0])
new_X=new_X.reshape((1,-1))
new_X=new_X.tolist()
else:
new_X=self.fitIGMM(obs,IsPlot)
# Test if x_max is repeated, if it is, draw another one at random
temp_new_X=[]
for idx,val in enumerate(new_X):
if np.all(np.any(np.abs(self.X-val)>0.02,axis=1)): # check if a data point is already taken
temp_new_X=np.append(temp_new_X,val)
if len(temp_new_X)==0:
temp_new_X=np.zeros((1,self.dim))
for idx in range(0,self.dim):
temp_new_X[0,idx]=np.random.uniform(self.scalebounds[idx,0],self.scalebounds[idx,1],1)
else:
temp_new_X=temp_new_X.reshape((-1,self.dim))
self.NumPoints=np.append(self.NumPoints,temp_new_X.shape[0])
finished_gmm_opt=time.time()
elapse_gmm_opt=finished_gmm_opt-start_gmm_opt
self.opt_time=np.hstack((self.opt_time,elapse_gmm_opt))
self.X=np.vstack((self.X, temp_new_X))
temp_X_new_original=[val*self.max_min_gap+self.bounds[:,0] for idx, val in enumerate(temp_new_X)]
temp_X_new_original=np.asarray(temp_X_new_original)
# Step 7 in the algorithm
# Evaluate y=f(x)
temp=self.f(temp_X_new_original)
temp=np.reshape(temp,(-1,1))
# Step 8 in the algorithm
self.Y=np.append(self.Y,temp)
self.X_original=np.vstack((self.X_original, temp_X_new_original))
print "#Batch={:d} f_max={:.3f}".format(temp_new_X.shape[0],self.Y.max())
def __init__(self,
gp_params,
f,
pbounds,
acq,
verbose=1,
opt_toolbox='scipy'):
"""
Input parameters
----------
f: function to optimize:
pbounds: bounds on parameters
acq: acquisition function, 'ei', 'ucb'
opt: optimization toolbox, 'nlopt','direct','scipy'
Returns
-------
dim: dimension
bounds: bounds on original scale
scalebounds: bounds on normalized scale of 0-1
time_opt: will record the time spent on optimization
gp: Gaussian Process object
"""
# Store the original dictionary
self.pbounds = pbounds
# Find number of parameters
self.dim = len(pbounds)
# Create an array with parameters bounds
if isinstance(pbounds, dict):
# Get the name of the parameters
self.keys = list(pbounds.keys())
self.bounds = []
for key in self.pbounds.keys():
self.bounds.append(self.pbounds[key])
self.bounds = np.asarray(self.bounds)
else:
self.bounds = np.asarray(pbounds)
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
# Some function to be optimized
self.f = f
# optimization tool: direct, scipy, nlopt
self.opt_toolbox = opt_toolbox
# acquisition function type
self.acq = acq
# store the batch size for each iteration
self.NumPoints = []
# Numpy array place holders
self.X_original = None
# scale the data to 0-1 fit GP better
self.X = None # X=( X_original - min(bounds) / (max(bounds) - min(bounds))
self.Y = None # Y=( Y_original - mean(bounds) / (max(bounds) - min(bounds))
self.Y_original = None
self.opt_time = 0
self.L = 0 # lipschitz
self.gp = PradaGaussianProcess(gp_params)
# Acquisition Function
#self.acq_func = None
self.acq_func = AcquisitionFunction(acq=self.acq)
def expandBoundsDDB_MAP(self):
"""
Description: Expands the search space with the MAP implementation of
our DDB method
"""
print('Attempting to expand search space with DDB-MAP method')
alpha=self.alpha
beta=self.beta
bound_samples=100 # Number of radius sample to fit the log-logistic distribution
# Find y^+ and x^+
ymax=np.max(self.Y)
# Generate test radii
max_loc=np.argmax(self.Y)
xmax=self.X[max_loc]
test_bound=np.zeros(self.scalebounds.shape)
bound_dist=np.zeros(bound_samples)
bound_center=xmax
test_bound[:,1]=bound_center+0.5
test_bound[:,0]=bound_center-0.5
max_radius=np.max(np.array([np.max(max_bound_size-test_bound[:,1]),np.max(test_bound[:,0])]))
step=max_radius/bound_samples
packing_number=np.zeros(bound_samples)
# Generate a Thompson sample maxima to estimate internal maxima
TS=AcquisitionFunction.ThompsonSampling(self.gp)
tsb_x,tsb_y=acq_max_global(TS, self.gp, bounds=self.scalebounds)
# Generate Gumbel samples to estimate the external maxima
for i in range(0,bound_samples):
bound_length=test_bound[:,1]-test_bound[:,0]
volume=np.power(max_bound_size,self.dim)-np.prod(bound_length)
packing_number[i]=round(volume/(5*self.gp.lengthscale))
mu=stats.norm.ppf(1.0-1.0/packing_number[i])
sigma=stats.norm.ppf(1.0-(1.0/packing_number[i])*np.exp(-1.0))-stats.norm.ppf(1.0-(1.0/(packing_number[i])))
bound_dist[i]=np.exp(-np.exp(-(-tsb_y-mu)/sigma))
test_bound[:,1]=test_bound[:,1]+step
test_bound[:,0]=test_bound[:,0]-step
bound_dist[np.isnan(bound_dist)]=1
# Fit the log-logistic paramaters to the Gumbel samples
xfit=np.arange(0,max_radius,max_radius/100)
popt,pcov=optimize.curve_fit(self.sufficientBoundPDF,xfit[0:100],bound_dist,bounds=np.array([[5,1.1],[20,5]]))
print("popt={}".format(popt))
b=ymax/popt[0]
a=popt[1]
print("b={}, ymax={}".format(b,ymax))
# Find the gamma and log-logistic modes to determine the optimisation bound
c=ymax/b
loglog_mode=a*np.power((c-1.0)/(c+1.0),(1/c))
gamma_mode=(alpha-1)/beta
opt_bound=np.ones([2])
opt_bound[0]=min(loglog_mode,gamma_mode)
opt_bound[1]=max(loglog_mode,gamma_mode)
bound_range=(opt_bound[1]-opt_bound[0])
# Find MAP Estimate of radius r
for d in range(0,self.dim):
r_max=0
p_max=0
for x0 in np.arange(opt_bound[0],opt_bound[1],bound_range/10):
res=optimize.minimize(lambda x: self.radiusPDF(x,alpha,beta,b,ymax,a),x0=x0, bounds=np.array([opt_bound]), method='L-BFGS-B')
if -res.fun>p_max:
r_max=res.x
p_max=-res.fun
if r_max>opt_bound[1]:
r_max=opt_bound[1]
xplot=np.arange(0,10,0.01)
yplot=-self.radiusPDF(xplot,alpha,beta,b,ymax,a)
max_loc=np.argmax(yplot)
print("optimal radius of {} with unscaled probability of {}".format(r_max,p_max))
self.scalebounds[d,1]=xmax[d]+r_max
self.scalebounds[d,0]=xmax[d]-r_max
print("seach space extended to {} with DDB".format(self.scalebounds))
class PradaBayOptFn(object):
def __init__(self,
gp_params,
f,
init_bounds,
pbounds,
acq,
verbose=1,
opt_toolbox='nlopt'):
"""
Input parameters
----------
f: function to optimize:
pbounds: bounds on parameters
acq: acquisition function, acq['name']=['ei','ucb','poi','lei']
,acq['kappa'] for ucb, acq['k'] for lei
opt: optimization toolbox, 'nlopt','direct','scipy'
Returns
-------
dim: dimension
bounds: bounds on original scale
scalebounds: bounds on normalized scale of 0-1
time_opt: will record the time spent on optimization
gp: Gaussian Process object
"""
# Find number of parameters
self.dim = len(pbounds)
# Create an array with parameters bounds
if isinstance(pbounds, dict):
# Get the name of the parameters
self.keys = list(pbounds.keys())
self.bounds = []
for key in pbounds.keys():
self.bounds.append(pbounds[key])
self.bounds = np.asarray(self.bounds)
else:
self.bounds = np.asarray(pbounds)
if len(init_bounds) == 0:
self.init_bounds = self.bounds.copy()
else:
self.init_bounds = init_bounds
if isinstance(init_bounds, dict):
# Get the name of the parameters
self.keys = list(init_bounds.keys())
self.init_bounds = []
for key in init_bounds.keys():
self.init_bounds.append(init_bounds[key])
self.init_bounds = np.asarray(self.init_bounds)
else:
self.init_bounds = np.asarray(init_bounds)
# create a scalebounds 0-1
scalebounds = np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds = scalebounds.T
self.max_min_gap = self.bounds[:, 1] - self.bounds[:, 0]
# Some function to be optimized
self.f = f
# optimization toolbox
self.opt_toolbox = opt_toolbox
# acquisition function type
self.acq = acq
# store X in original scale
self.X_original = None
# store X in 0-1 scale
self.X = None
# store y=f(x)
# (y - mean)/(max-min)
self.Y = None
# y original scale
self.Y_original = None
self.time_opt = 0
self.k_Neighbor = 2
# Lipschitz constant
self.L = 0
# Gaussian Process class
self.gp = PradaGaussianProcess(gp_params)
# acquisition function
self.acq_func = None
# stop condition
self.stop_flag = 0
# will be later used for visualization
def posterior(self, Xnew):
self.gp.fit(self.X, self.Y)
mu, sigma2 = self.gp.predict(Xnew, eval_MSE=True)
return mu, np.sqrt(sigma2)
def init(self, gp_params, n_init_points=3):
"""
Input parameters
----------
gp_params: Gaussian Process structure
n_init_points: # init points
"""
# Generate random points
l = [
np.random.uniform(x[0], x[1], size=n_init_points)
for x in self.init_bounds
]
# Concatenate new random points to possible existing
# points from self.explore method.
temp = np.asarray(l)
temp = temp.T
init_X = list(temp.reshape((n_init_points, -1)))
self.X_original = np.asarray(init_X)
# Evaluate target function at all initialization
y_init = self.f(init_X)
y_init = np.reshape(y_init, (n_init_points, 1))
self.Y_original = np.asarray(y_init)
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
# convert it to scaleX
temp_init_point = np.divide((init_X - self.bounds[:, 0]),
self.max_min_gap)
self.X = np.asarray(temp_init_point)
def estimate_L(self, bounds):
'''
Estimate the Lipschitz constant of f by taking maximizing the norm of the expectation of the gradient of *f*.
'''
def df(x, model, x0):
mean_derivative = gp_model.predictive_gradient(self.X, self.Y, x)
temp = mean_derivative * mean_derivative
if len(temp.shape) <= 1:
res = np.sqrt(temp)
else:
res = np.sqrt(
np.sum(temp, axis=1)
) # simply take the norm of the expectation of the gradient
return -res
gp_model = self.gp
dim = len(bounds)
num_data = 1000 * dim
samples = np.zeros(shape=(num_data, dim))
for k in range(0, dim):
samples[:, k] = np.random.uniform(low=bounds[k][0],
high=bounds[k][1],
size=num_data)
#samples = np.vstack([samples,gp_model.X])
pred_samples = df(samples, gp_model, 0)
x0 = samples[np.argmin(pred_samples)]
res = minimize(df,
x0,
method='L-BFGS-B',
bounds=bounds,
args=(gp_model, x0),
options={'maxiter': 100})
try:
minusL = res.fun[0][0]
except:
if len(res.fun.shape) == 1:
minusL = res.fun[0]
else:
minusL = res.fun
L = -minusL
if L < 1e-6:
L = 0.0001 ## to avoid problems in cases in which the model is flat.
return L
def maximize(self, gp_params, kappa=2):
"""
Main optimization method.
Input parameters
----------
kappa: parameter for UCB acquisition only.
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
acq = self.acq
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
if acq['name'] == 'nei':
self.L = self.estimate_L(self.scalebounds)
self.acq_func = AcquisitionFunction(kind=self.acq, L=self.L)
else:
self.acq_func = AcquisitionFunction(self.acq)
if acq['name'] == "ei_mu":
#find the maximum in the predictive mean
mu_acq = {}
mu_acq['name'] = 'mu'
mu_acq['dim'] = self.dim
acq_mu = AcquisitionFunction(mu_acq)
x_mu_max = acq_max(ac=acq_mu.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
# set y_max = mu_max
y_max = acq_mu.acq_kind(x_mu_max, gp=self.gp, y_max=y_max)
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
val_acq = self.acq_func.acq_kind(x_max, self.gp, y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# check the value alpha(x_max)==0
#if val_acq<0.0001:
#self.stop_flag=1
#return
# select the optimization toolbox
"""
if self.opt=='nlopt':
x_max,f_max = acq_max_nlopt(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
if self.opt=='scipy':
if self.opt=='direct':
x_max = acq_max_direct(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
"""
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max).sum(axis=1) == 0):
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# store X
self.X = np.vstack((self.X, x_max.reshape((1, -1))))
# compute X in original scale
temp_X_new_original = x_max * self.max_min_gap + self.bounds[:, 0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
def maximize(self, gp_params, kappa=2):
"""
Main optimization method.
Input parameters
----------
kappa: parameter for UCB acquisition only.
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=1) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.time_opt = np.hstack((self.time_opt, 0))
return
# init a new Gaussian Process
self.gp = PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Set acquisition function
start_opt = time.time()
acq = self.acq
y_max = self.Y.max()
#self.L=self.estimate_L(self.scalebounds)
# select the acquisition function
if acq['name'] == 'nei':
self.L = self.estimate_L(self.scalebounds)
self.acq_func = AcquisitionFunction(kind=self.acq, L=self.L)
else:
self.acq_func = AcquisitionFunction(self.acq)
if acq['name'] == "ei_mu":
#find the maximum in the predictive mean
mu_acq = {}
mu_acq['name'] = 'mu'
mu_acq['dim'] = self.dim
acq_mu = AcquisitionFunction(mu_acq)
x_mu_max = acq_max(ac=acq_mu.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
# set y_max = mu_max
y_max = acq_mu.acq_kind(x_mu_max, gp=self.gp, y_max=y_max)
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.scalebounds,
opt_toolbox=self.opt_toolbox)
val_acq = self.acq_func.acq_kind(x_max, self.gp, y_max)
#print "alpha[x_max]={:.5f}".format(np.ravel(val_acq)[0])
# check the value alpha(x_max)==0
#if val_acq<0.0001:
#self.stop_flag=1
#return
# select the optimization toolbox
"""
if self.opt=='nlopt':
x_max,f_max = acq_max_nlopt(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
if self.opt=='scipy':
if self.opt=='direct':
x_max = acq_max_direct(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
"""
# record the optimization time
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.time_opt = np.hstack((self.time_opt, elapse_opt))
# Test if x_max is repeated, if it is, draw another one at random
if np.any((self.X - x_max).sum(axis=1) == 0):
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
# store X
self.X = np.vstack((self.X, x_max.reshape((1, -1))))
# compute X in original scale
temp_X_new_original = x_max * self.max_min_gap + self.bounds[:, 0]
self.X_original = np.vstack((self.X_original, temp_X_new_original))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original,
self.f(temp_X_new_original))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
def reduced_eval_consistent_bayesian_model(
bayesian_model: mc_dropout.BayesianModule,
acquisition_function: AcquisitionFunction,
num_classes: int,
k: int,
initial_percentage: int,
reduce_percentage: int,
target_size: int,
available_loader,
device=None,
) -> SubsetEvalResults:
"""Performs a scoring step with k inference samples while reducing the dataset to at most min_remaining_percentage.
Before computing anything at all the initial available dataset is randomly culled to initial_percentage.
Every `chunk_size` inferences BALD is recomputed and the bottom `reduce_percentage` samples are dropped."""
global reduced_eval_consistent_bayesian_model_cuda_chunk_size
# TODO: ActiveLearningData should be renamed to be a more modular SplitDataset.
# Here, we need to use available_dataset because it allows us to easily recover the original indices.
# We start with all data in the acquired data.
subset_split = active_learning_data.ActiveLearningData(
available_loader.dataset)
initial_length = len(available_loader.dataset)
initial_split_length = initial_length * initial_percentage // 100
# By acquiring [initial_split_length:], we make the tail unavailable.
subset_split.acquire(torch.randperm(initial_length)[initial_split_length:])
subset_dataloader = data.DataLoader(subset_split.available_dataset,
shuffle=False,
batch_size=available_loader.batch_size)
print(f"Scoring subset of {len(subset_dataloader.dataset)} items:")
# We're done with available_loader in this function.
available_loader = None
with torch.no_grad():
B = len(subset_split.available_dataset)
C = num_classes
# We stay on the CPU.
logits_B_K_C = None
k_lower = 0
torch_utils.gc_cuda()
chunk_size = reduced_eval_consistent_bayesian_model_cuda_chunk_size if device.type == "cuda" else 32
while k_lower < k:
try:
k_upper = min(k_lower + chunk_size, k)
old_logit_B_K_C = logits_B_K_C
# This also stays on the CPU.
logits_B_K_C = torch.empty((B, k_upper, C),
dtype=torch.float64)
# Copy the old data over.
if k_lower > 0:
logits_B_K_C[:, 0:k_lower, :].copy_(old_logit_B_K_C)
old_logit_B_K_C = None
# This resets the dropout masks.
bayesian_model.eval()
for i, (batch, _) in enumerate(
with_progress_bar(
subset_dataloader,
unit_scale=subset_dataloader.batch_size)):
lower = i * subset_dataloader.batch_size
upper = min(lower + subset_dataloader.batch_size, B)
batch = batch.to(device)
# batch_size x ws x classes
mc_output_B_K_C = bayesian_model(batch, k_upper - k_lower)
logits_B_K_C[lower:upper, k_lower:k_upper].copy_(
mc_output_B_K_C.double(), non_blocking=True)
except RuntimeError as exception:
if torch_utils.should_reduce_batch_size(exception):
if chunk_size <= 1:
raise
chunk_size = chunk_size // 2
print(
f"New reduced_eval_consistent_bayesian_model_cuda_chunk_size={chunk_size} ({exception})"
)
reduced_eval_consistent_bayesian_model_cuda_chunk_size = chunk_size
torch_utils.gc_cuda()
else:
raise
else:
if k_upper == k:
next_size = target_size
elif k_upper < 50:
next_size = B
else:
next_size = max(target_size,
B * (100 - reduce_percentage) // 100)
# Compute the score if it's needed: we are going to reduce the dataset or we're in the last iteration.
if next_size < B or k_upper == k:
# Calculate the scores (mutual information) of logits_B_K_C
scores_B = acquisition_function.compute_scores(
logits_B_K_C,
available_loader=subset_dataloader,
device=device)
else:
scores_B = None
if next_size < B:
print("Reducing size", next_size)
# Get indices of samples sorted by increasing mutual information
sorted_indices = torch.argsort(scores_B, descending=True)
# Select next_size samples with smallest mutual information (ascending indices)
new_indices = torch.sort(sorted_indices[:next_size],
descending=False)[0]
B = next_size
logits_B_K_C = logits_B_K_C[new_indices]
if k_upper == k:
logits_B_K_C = logits_B_K_C.clone().detach()
scores_B = scores_B[new_indices].clone().detach()
# Acquire all the low scorers
subset_split.acquire(sorted_indices[next_size:])
k_lower += chunk_size
return SubsetEvalResults(subset_split=subset_split,
subset_dataloader=subset_dataloader,
scores_B=scores_B,
logits_B_K_C=logits_B_K_C)
def maximize_batch_CL_incremental(self, gp_params, B=5):
"""
Finding a batch of points using Constant Liar approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
kappa: constant value in UCB
Returns
-------
X: a batch of [x_1..x_Nt]
"""
self.NumPoints = np.append(self.NumPoints, B)
if self.acq['name'] == 'random':
x_max = [
np.random.uniform(x[0], x[1], size=B) for x in self.bounds
]
x_max = np.asarray(x_max)
x_max = x_max.T
self.X_original = np.vstack((self.X_original, x_max))
# evaluate Y using original X
#self.Y = np.append(self.Y, self.f(temp_X_new_original))
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
self.opt_time = np.hstack((self.opt_time, 0))
return
#const_liar=self.Y.mean()
#const_liar=self.Y_original.min()
#const_liar=self.Y.max()
# Set acquisition function
self.acq_func = AcquisitionFunction(self.acq)
y_max = self.Y.max()
# Set parameters if any was passed
self.gp = PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
start_opt = time.time()
# copy GP, X and Y
temp_gp = copy.deepcopy(self.gp)
temp_X = self.X
temp_Y = self.Y
#temp_Y_original=self.Y_original
#store new_x
new_X = []
for ii in range(B):
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=temp_gp,
y_max=y_max,
bounds=self.scalebounds)
# Test if x_max is repeated, if it is, draw another one at random
if np.any(
np.any(np.abs(self.X - x_max) < 0.02,
axis=1)): # check if a data point is already taken
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
if ii == 0:
new_X = x_max
else:
new_X = np.vstack((new_X, x_max.reshape((1, -1))))
const_liar = temp_gp.predict(x_max, eval_MSE=true)
#temp_X= np.vstack((temp_X, x_max.reshape((1, -1))))
#temp_Y = np.append(temp_Y, const_liar )
#temp_gp.fit(temp_X,temp_Y)
# update the Gaussian Process and thus the acquisition function
#temp_gp.compute_incremental_var(temp_X,x_max)
temp_gp.fit_incremental(x_max, np.asarray([const_liar]))
# Updating the GP.
new_X = new_X.reshape((B, -1))
finished_opt = time.time()
elapse_opt = finished_opt - start_opt
self.opt_time = np.hstack((self.opt_time, elapse_opt))
#print new_X
self.X = np.vstack((self.X, new_X))
# convert back to original scale
temp_X_new_original = [
val * self.max_min_gap + self.bounds[:, 0]
for idx, val in enumerate(new_X)
]
temp_X_new_original = np.asarray(temp_X_new_original)
self.X_original = np.vstack((self.X_original, temp_X_new_original))
for idx, val in enumerate(temp_X_new_original):
self.Y_original = np.append(self.Y_original, self.f(val))
# update Y after change Y_original
self.Y = (self.Y_original - np.mean(self.Y_original)) / (
np.max(self.Y_original) - np.min(self.Y_original))
def maximize_batch_BUCB(self,gp_params, B=5,kappa=2):
"""
Finding a batch of points using GP-BUCB approach
Input Parameters
----------
gp_params: Parameters to be passed to the Gaussian Process class
B: fixed batch size for all iteration
kappa: constant value in UCB
IsPlot: flag variable for visualization
Returns
-------
X: a batch of [x_1..x_B]
"""
self.B=B
# Set acquisition function
self.acq_func = AcquisitionFunction(kind=self.acq, kappa=kappa)
# Set parameters if any was passed
self.gp=PradaGaussianProcess(gp_params)
if len(self.gp.KK_x_x_inv)==0: # check if empty
self.gp.fit(self.X, self.Y)
#else:
#self.gp.fit_incremental(self.X[ur], self.Y[ur])
start_gmm_opt=time.time()
# generate samples from Acquisition function
y_max=self.gp.Y.max()
# check the bound 0-1 or original bound
temp_X=self.X
temp_gp=self.gp
temp_gp.X_bucb=temp_X
temp_gp.KK_x_x_inv_bucb=self.gp.KK_x_x_inv
# finding new X
new_X=[]
for ii in range(B):
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind, gp=temp_gp, y_max=y_max, bounds=self.scalebounds)
if np.any((temp_X - x_max).sum(axis=1) == 0) | np.isnan(x_max.sum()):
x_max = np.random.uniform(self.scalebounds[:, 0],
self.scalebounds[:, 1],
size=self.scalebounds.shape[0])
if ii==0:
new_X=x_max
else:
new_X= np.vstack((new_X, x_max.reshape((1, -1))))
# update the Gaussian Process and thus the acquisition function
temp_gp.compute_incremental_var(temp_X,x_max)
temp_X = np.vstack((temp_X, x_max.reshape((1, -1))))
temp_gp.X_bucb=temp_X
# record the optimization time
finished_gmm_opt=time.time()
elapse_gmm_opt=finished_gmm_opt-start_gmm_opt
self.opt_time=np.hstack((self.opt_time,elapse_gmm_opt))
self.NumPoints=np.append(self.NumPoints,B)
self.X=temp_X
# convert back to original scale
temp_X_new_original=[val*self.max_min_gap+self.bounds[:,0] for idx, val in enumerate(new_X)]
temp_X_new_original=np.asarray(temp_X_new_original)
self.X_original=np.vstack((self.X_original, temp_X_new_original))
# evaluate y=f(x)
temp=self.f(temp_X_new_original)
temp=np.reshape(temp,(-1,1))
self.Y=np.append(self.Y,temp)
print "#Batch={:d} f_max={:.4f}".format(new_X.shape[0],self.Y.max())
def maximize(self,gp_params):
"""
Main optimization method.
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.stop_flag==1:
return
if self.acq['name']=='random':
x_max = [np.random.uniform(x[0], x[1], size=1) for x in self.scalebounds]
x_max=np.asarray(x_max)
x_max=x_max.T
self.X_original=np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y=(self.Y_original-np.mean(self.Y_original))/np.std(self.Y_original)
self.time_opt=np.hstack((self.time_opt,0))
return
# init a new Gaussian Process
self.gp=PradaGaussianProcess(gp_params)
if self.gp.KK_x_x_inv ==[]:
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
acq=self.acq
if acq['debug']==1:
logmarginal=self.gp.log_marginal_lengthscale(gp_params['theta'],gp_params['noise_delta'])
print(gp_params['theta'])
print("log marginal before optimizing ={:.4f}".format(logmarginal))
self.logmarginal=logmarginal
if logmarginal<-999999:
logmarginal=self.gp.log_marginal_lengthscale(gp_params['theta'],gp_params['noise_delta'])
if self.optimize_gp==1 and len(self.Y)%2*self.dim==0 and len(self.Y)>5*self.dim:
print("Initial length scale={}".format(gp_params['theta']))
newtheta = self.gp.optimize_lengthscale(gp_params['theta'],gp_params['noise_delta'],self.scalebounds)
gp_params['theta']=newtheta
print("New length scale={}".format(gp_params['theta']))
# init a new Gaussian Process after optimizing hyper-parameter
self.gp=PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Modify search space based on selected method
if self.expandSS=='expandBoundsDDB_MAP':
self.expandBoundsDDB_MAP()
if self.expandSS=='expandBoundsDDB_FB':
self.expandBoundsDDB_FB()
if self.expandSS=='expandBoundsFiltering':
self.expandBoundsFiltering()
if self.expandSS=='volumeDoubling' and len(self.Y)%3*self.dim==0:
self.volumeDoubling()
# Prevent bounds from breaching maximum limit
for d in range(0,self.dim):
if self.scalebounds[d,0]<0:
print('Lower bound of {} in dimention {} exceeded minimum bound of {}. Scaling up.'.format(self.scalebounds[d,0],d,0))
self.scalebounds[d,0]=0
print('bound set to {}'.format(self.scalebounds))
if self.scalebounds[d,1]>max_bound_size:
print('Upper bound of {} in dimention {} exceeded maximum bound of {}. Scaling down.'.format(self.scalebounds[d,1],d,max_bound_size))
self.scalebounds[d,1]=max_bound_size
self.scalebounds[d,0]=min(self.scalebounds[d,0],self.scalebounds[d,1]-np.sqrt(3*self.gp.lengthscale))
print('bound set to {}'.format(self.scalebounds))
# Set acquisition function
start_opt=time.time()
y_max = self.Y.max()
if acq['name'] in ['consensus','mes']:
ucb_acq_func={}
ucb_acq_func['name']='ucb'
ucb_acq_func['kappa']=np.log(len(self.Y))
ucb_acq_func['dim']=self.dim
ucb_acq_func['scalebounds']=self.scalebounds
myacq=AcquisitionFunction(ucb_acq_func)
xt_ucb = acq_max(ac=myacq.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
xstars=[]
xstars.append(xt_ucb)
ei_acq_func={}
ei_acq_func['name']='ei'
ei_acq_func['dim']=self.dim
ei_acq_func['scalebounds']=self.scalebounds
myacq=AcquisitionFunction(ei_acq_func)
xt_ei = acq_max(ac=myacq.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
xstars.append(xt_ei)
pes_acq_func={}
pes_acq_func['name']='pes'
pes_acq_func['dim']=self.dim
pes_acq_func['scalebounds']=self.scalebounds
myacq=AcquisitionFunction(pes_acq_func)
xt_pes = acq_max(ac=myacq.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
xstars.append(xt_pes)
self.xstars=xstars
if acq['name']=='vrs':
print("please call the maximize_vrs function")
return
if 'xstars' not in globals():
xstars=[]
self.xstars=xstars
self.acq['xstars']=xstars
self.acq['WW']=False
self.acq['WW_dim']=False
self.acq_func = AcquisitionFunction(self.acq,self.bb_function)
if acq['name']=="ei_mu":
#find the maximum in the predictive mean
mu_acq={}
mu_acq['name']='mu'
mu_acq['dim']=self.dim
acq_mu=AcquisitionFunction(mu_acq)
x_mu_max = acq_max(ac=acq_mu.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds,opt_toolbox=self.opt_toolbox)
# set y_max = mu_max
y_max=acq_mu.acq_kind(x_mu_max,gp=self.gp, y_max=y_max)
x_max = acq_max(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds,opt_toolbox=self.opt_toolbox,seeds=self.xstars)
if acq['name']=='consensus' and acq['debug']==1: # plot the x_max and xstars
fig=plt.figure(figsize=(5, 5))
plt.scatter(xt_ucb[0],xt_ucb[1],marker='s',color='g',s=200,label='Peak')
plt.scatter(xt_ei[0],xt_ei[1],marker='s',color='k',s=200,label='Peak')
plt.scatter(x_max[0],x_max[1],marker='*',color='r',s=300,label='Peak')
plt.xlim(0,1)
plt.ylim(0,1)
strFileName="acquisition_functions_debug.eps"
fig.savefig(strFileName, bbox_inches='tight')
if acq['name']=='vrs' and acq['debug']==1: # plot the x_max and xstars
fig=plt.figure(figsize=(5, 5))
plt.scatter(xt_ucb[0],xt_ucb[1],marker='s',color='g',s=200,label='Peak')
plt.scatter(xt_ei[0],xt_ei[1],marker='s',color='k',s=200,label='Peak')
plt.scatter(x_max[0],x_max[1],marker='*',color='r',s=300,label='Peak')
plt.xlim(0,1)
plt.ylim(0,1)
strFileName="vrs_acquisition_functions_debug.eps"
#fig.savefig(strFileName, bbox_inches='tight')
val_acq=self.acq_func.acq_kind(x_max,self.gp,y_max)
#print x_max
#print val_acq
if self.stopping_criteria!=0 and val_acq<self.stopping_criteria:
val_acq=self.acq_func.acq_kind(x_max,self.gp,y_max)
self.stop_flag=1
print("Stopping Criteria is violated. Stopping Criteria is {:.15f}".format(self.stopping_criteria))
self.alpha_Xt= np.append(self.alpha_Xt,val_acq)
mean,var=self.gp.predict(x_max, eval_MSE=True)
var.flags['WRITEABLE']=True
var[var<1e-20]=0
#self.Tau_Xt= np.append(self.Tau_Xt,val_acq/var)
# record the optimization time
finished_opt=time.time()
elapse_opt=finished_opt-start_opt
self.time_opt=np.hstack((self.time_opt,elapse_opt))
# store X
self.X = np.vstack((self.X, x_max.reshape((1, -1))))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y=(self.Y_original-np.mean(self.Y_original))/np.std(self.Y_original)
if self.gp.flagIncremental==1:
self.gp.fit_incremental(x_max,self.Y[-1])
# if (self.acq['name']=='ei_regularizerH') or (self.acq['name']=='ei_regularizerQ'):
# self.scalebounds[:,0]=self.scalebounds[:,0]+1
# self.scalebounds[:,1]=self.scalebounds[:,1]-1
# self.acq['scalebounds']=self.scalebounds
self.experiment_num=self.experiment_num+1
class PradaBayOptFn(object):
def __init__(self, gp_params, func_params, acq_params, experiment_num, seed):
"""
Input parameters
----------
gp_params: GP parameters
gp_params.theta: to compute the kernel
gp_params.delta: to compute the kernel
func_params: function to optimize
func_params.init bound: initial bounds for parameters
func_params.bounds: bounds on parameters
func_params.func: a function to be optimized
acq_params: acquisition function,
acq_params.acq_func['name']=['ei','ucb','poi','lei']
,acq['kappa'] for ucb, acq['k'] for lei
acq_params.opt_toolbox: optimization toolbox 'nlopt','direct','scipy'
experiment_num: the interation of the GP method. Used to make sure each
independant stage of the experiment uses different
initial conditions
seed: Variable used as part of a seed to generate random initial points
Returns
-------
dim: dimension
scalebounds: bound used thoughout the BO algorithm
time_opt: will record the time spent on optimization
gp: Gaussian Process object
"""
self.experiment_num=experiment_num
self.seed=seed
np.random.seed(self.experiment_num*self.seed)
# Prior distribution paramaters for the DDB method
self.alpha=2
self.beta=4
# Find number of parameters
bounds=func_params['bounds']
if 'init_bounds' not in func_params:
init_bounds=bounds
else:
init_bounds=func_params['init_bounds']
# Find input dimention
self.dim = len(bounds)
self.radius=np.ones([self.dim,1])
# Generate bound array
scalebounds=np.array([np.zeros(self.dim), np.ones(self.dim)])
self.scalebounds=scalebounds.T
# find function to be optimized
self.f = func_params['f']
# acquisition function type
self.acq=acq_params['acq_func']
# Check if the search space is to be modified
self.bb_function=acq_params["bb_function"]
if 'expandSS' not in acq_params:
self.expandSS=0
else:
self.expandSS=acq_params['expandSS']
# Check if the bound is to be set randomly. If so, shift the bound by a random amount
if (acq_params['random_initial_bound']==1):
randomizer=np.random.rand(self.dim)*max_bound_size
for d in range(0,self.dim):
self.scalebounds[d]=self.scalebounds[d]+randomizer[d]
# Other checks
if 'debug' not in self.acq:
self.acq['debug']=0
if 'stopping' not in acq_params:
self.stopping_criteria=0
else:
self.stopping_criteria=acq_params['stopping']
if 'optimize_gp' not in acq_params:
self.optimize_gp=0
else:
self.optimize_gp=acq_params['optimize_gp']
if 'marginalize_gp' not in acq_params:
self.marginalize_gp=0
else:
self.marginalize_gp=acq_params['marginalize_gp']
# optimization toolbox
if 'opt_toolbox' not in acq_params:
if self.acq['name']=='ei_reg':
self.opt_toolbox='unbounded'
else:
self.opt_toolbox='scipy'
else:
self.opt_toolbox=acq_params['opt_toolbox']
self.iteration_factor=acq_params['iteration_factor']
# store X in original scale
self.X_original= None
# store X in 0-1 scale
self.X = None
# store y=f(x)
# (y - mean)/(max-min)
self.Y = None
# y original scale
self.Y_original = None
# value of the acquisition function at the selected point
self.alpha_Xt=None
self.Tau_Xt=None
self.time_opt=0
self.k_Neighbor=2
# Gaussian Process class
self.gp=PradaGaussianProcess(gp_params)
self.gp_params=gp_params
# acquisition function
self.acq_func = None
# stop condition
self.stop_flag=0
self.logmarginal=0
# xt_suggestion, caching for Consensus
self.xstars=[]
self.ystars=np.zeros((2,1))
# theta vector for marginalization GP
self.theta_vector =[]
def init(self,gp_params, n_init_points=3):
"""
Input parameters
----------
gp_params: Gaussian Process structure
n_init_points: # init points
"""
# set seed to allow for reproducible results
np.random.seed(self.experiment_num*self.seed)
print(self.experiment_num)
#Generate initial points on grid
l=np.zeros([n_init_points,self.dim])
bound_length=self.scalebounds[0,1]-self.scalebounds[0,0]
for d in range(0,self.dim):
l[:,d]=lhs(n_init_points)[:,0]
self.X=np.asarray(l)+self.scalebounds[:,0]
self.X=self.X*bound_length #initial inouts
print("starting points={}".format(self.X))
print("starting bounds={}".format(self.scalebounds))
y_init=self.f(self.X)
y_init=np.reshape(y_init,(n_init_points,1))
self.Y_original = np.asarray(y_init) #initial outputs
self.Y=(self.Y_original-np.mean(self.Y_original))/np.std(self.Y_original) #outputs normalised
print("starting Y values={}".format(self.Y))
#############Rename#############
def radiusPDF(self,r,alpha,beta,b,ymax,a):
"""
Description: Evaluates the posterior distribution for our DDB method
Input parameters
----------
r: radius to be evaluated
alpha: # gamma distribution shape paramater
beta: # gamma distribution rate paramater
a: # log-logistic distribution scale paramater
b: # log-logistic distribution rate paramater with y_max
y_max: # log-logistic distribution rate paramater with b
Output: posterior distribution evaluated at r
"""
gamma=stats.gamma.pdf(r,alpha,scale=1/beta)
loglog=stats.fisk.pdf(r,ymax/b,scale=a)
P=gamma*loglog
return -P
def sufficientBoundPDF(self,r,bDivYmax,a):
"""
Description: Evaluates the likelihood distribution for our DDB method
Input parameters
----------
r: radius to be evaluated
a: # log-logistic distribution scale paramater
bDivYmax: # log-logistic distribution rate paramater
Output: likelihood distribution evaluated at r
"""
P=stats.fisk.cdf(r,bDivYmax,scale=a)
return P
def expandBoundsDDB_MAP(self):
"""
Description: Expands the search space with the MAP implementation of
our DDB method
"""
print('Attempting to expand search space with DDB-MAP method')
alpha=self.alpha
beta=self.beta
bound_samples=100 # Number of radius sample to fit the log-logistic distribution
# Find y^+ and x^+
ymax=np.max(self.Y)
# Generate test radii
max_loc=np.argmax(self.Y)
xmax=self.X[max_loc]
test_bound=np.zeros(self.scalebounds.shape)
bound_dist=np.zeros(bound_samples)
bound_center=xmax
test_bound[:,1]=bound_center+0.5
test_bound[:,0]=bound_center-0.5
max_radius=np.max(np.array([np.max(max_bound_size-test_bound[:,1]),np.max(test_bound[:,0])]))
step=max_radius/bound_samples
packing_number=np.zeros(bound_samples)
# Generate a Thompson sample maxima to estimate internal maxima
TS=AcquisitionFunction.ThompsonSampling(self.gp)
tsb_x,tsb_y=acq_max_global(TS, self.gp, bounds=self.scalebounds)
# Generate Gumbel samples to estimate the external maxima
for i in range(0,bound_samples):
bound_length=test_bound[:,1]-test_bound[:,0]
volume=np.power(max_bound_size,self.dim)-np.prod(bound_length)
packing_number[i]=round(volume/(5*self.gp.lengthscale))
mu=stats.norm.ppf(1.0-1.0/packing_number[i])
sigma=stats.norm.ppf(1.0-(1.0/packing_number[i])*np.exp(-1.0))-stats.norm.ppf(1.0-(1.0/(packing_number[i])))
bound_dist[i]=np.exp(-np.exp(-(-tsb_y-mu)/sigma))
test_bound[:,1]=test_bound[:,1]+step
test_bound[:,0]=test_bound[:,0]-step
bound_dist[np.isnan(bound_dist)]=1
# Fit the log-logistic paramaters to the Gumbel samples
xfit=np.arange(0,max_radius,max_radius/100)
popt,pcov=optimize.curve_fit(self.sufficientBoundPDF,xfit[0:100],bound_dist,bounds=np.array([[5,1.1],[20,5]]))
print("popt={}".format(popt))
b=ymax/popt[0]
a=popt[1]
print("b={}, ymax={}".format(b,ymax))
# Find the gamma and log-logistic modes to determine the optimisation bound
c=ymax/b
loglog_mode=a*np.power((c-1.0)/(c+1.0),(1/c))
gamma_mode=(alpha-1)/beta
opt_bound=np.ones([2])
opt_bound[0]=min(loglog_mode,gamma_mode)
opt_bound[1]=max(loglog_mode,gamma_mode)
bound_range=(opt_bound[1]-opt_bound[0])
# Find MAP Estimate of radius r
for d in range(0,self.dim):
r_max=0
p_max=0
for x0 in np.arange(opt_bound[0],opt_bound[1],bound_range/10):
res=optimize.minimize(lambda x: self.radiusPDF(x,alpha,beta,b,ymax,a),x0=x0, bounds=np.array([opt_bound]), method='L-BFGS-B')
if -res.fun>p_max:
r_max=res.x
p_max=-res.fun
if r_max>opt_bound[1]:
r_max=opt_bound[1]
xplot=np.arange(0,10,0.01)
yplot=-self.radiusPDF(xplot,alpha,beta,b,ymax,a)
max_loc=np.argmax(yplot)
print("optimal radius of {} with unscaled probability of {}".format(r_max,p_max))
self.scalebounds[d,1]=xmax[d]+r_max
self.scalebounds[d,0]=xmax[d]-r_max
print("seach space extended to {} with DDB".format(self.scalebounds))
def expandBoundsDDB_FB(self):
"""
Description: Expands the search space with the full Bayesian
implementation of our DDB method
"""
print('Attempting to expand search space with DDB-FB method')
alpha=self.alpha
beta=self.beta
bound_samples=100 # Number of radius sample to fit the log-logistic distribution
# Find y^+ and x^+
ymax=np.max(self.Y)
# Generate test radii
max_loc=np.argmax(self.Y)
xmax=self.X[max_loc]
test_bound=np.zeros(self.scalebounds.shape)
bound_dist=np.zeros(bound_samples)
bound_center=xmax
test_bound[:,1]=bound_center+0.5
test_bound[:,0]=bound_center-0.5
max_radius=np.max(np.array([np.max(max_bound_size-test_bound[:,1]),np.max(test_bound[:,0])]))
step=max_radius/bound_samples
packing_number=np.zeros(bound_samples)
# Generate a Thompson sample maxima to estimate internal maxima
TS=AcquisitionFunction.ThompsonSampling(self.gp)
tsb_x,tsb_y=acq_max_global(TS, self.gp, bounds=self.scalebounds)
# Generate Gumbel samples to estimate the external maxima
for i in range(0,bound_samples):
bound_length=test_bound[:,1]-test_bound[:,0]
volume=np.power(max_bound_size,self.dim)-np.prod(bound_length)
packing_number[i]=round(volume/(5*self.gp.lengthscale))
mu=stats.norm.ppf(1.0-1.0/packing_number[i])
sigma=stats.norm.ppf(1.0-(1.0/packing_number[i])*np.exp(-1.0))-stats.norm.ppf(1.0-(1.0/(packing_number[i])))
bound_dist[i]=np.exp(-np.exp(-(-tsb_y-mu)/sigma))
test_bound[:,1]=test_bound[:,1]+step
test_bound[:,0]=test_bound[:,0]-step
bound_dist[np.isnan(bound_dist)]=1
# Fit the log-logistic paramaters to the Gumbel samples
xfit=np.arange(0,max_radius,max_radius/100)
popt,pcov=optimize.curve_fit(self.sufficientBoundPDF,xfit[0:100],bound_dist,bounds=np.array([[5,1.1],[20,5]]))
print("popt={}".format(popt))
b=ymax/popt[0]
a=popt[1]
print("b={}, ymax={}".format(b,ymax))
# Sample for the optimal radius
for d in range(0,self.dim):
gamma=np.random.gamma(shape=alpha,scale=1/beta,size=100)
loglog=stats.fisk.pdf(gamma,ymax/b,scale=a)
scaled_weights=loglog/np.sum(loglog)
multi=np.random.multinomial(1,scaled_weights)
r_index=np.argmax(multi)
print("Radius of {} selected".format(gamma[r_index]))
self.scalebounds[d,1]=xmax[d]+gamma[r_index]
self.scalebounds[d,0]=xmax[d]-gamma[r_index]
print("seach space extended to {} with DDB".format(self.scalebounds))
def lcb(self,x, gp):
"""
Calculates the GP-LCB acquisition function values
Inputs: gp: The Gaussian process, also contains all data
x:The point at which to evaluate the acquisition function
Output: acq_value: The value of the aquisition function at point x
"""
mean, var = gp.predict(x, eval_MSE=True)
var.flags['WRITEABLE']=True
var[var<1e-10]=0 #prevents negative variances obtained through comp errors
mean=np.atleast_2d(mean).T
var=np.atleast_2d(var).T
beta=2*np.log(len(gp.Y)*np.square((self.experiment_num+1)*math.pi)/(6*0.9))
return mean - np.sqrt(beta) * np.sqrt(var)
def ucb(self,x, gp):
"""
Calculates the GP-UCB acquisition function values
Inputs: gp: The Gaussian process, also contains all data
x:The point at which to evaluate the acquisition function
Output: acq_value: The value of the aquisition function at point x
"""
mean, var = gp.predict(x, eval_MSE=True)
var.flags['WRITEABLE']=True
var[var<1e-10]=0 #prevents negative variances obtained through comp errors
mean=np.atleast_2d(mean).T
var=np.atleast_2d(var).T
beta=2*np.log(len(gp.Y)*np.square(self.experiment_num*math.pi)/(6*0.9))
return mean + np.sqrt(beta) * np.sqrt(var)
def expandBoundsFiltering(self):
"""
Description: Expands the search space with filtering Bayesian
optimisation (FBO) by Nguyen et al.
"""
step=0.1*self.gp.lengthscale
print('Attempting to expand search space with FBO method')
# Determine the unfiltered extension based on the iteration number
extended_bound=np.copy(self.scalebounds)
extention=math.pow(self.iteration_factor/(max([self.experiment_num,1])),(1/self.dim))
old_radius=(extended_bound[:,1]-extended_bound[:,0])/2
mid_point=extended_bound[:,0]+old_radius
new_radius=old_radius*extention
extended_bound[:,1]=mid_point+new_radius
extended_bound[:,0]=mid_point-new_radius
# Calculate the global maximum lower confidence bound
lcb_x,lcb_y=acq_max_global(self.lcb, self.gp, extended_bound)
# Filter the lower and upper boundary up to the unfiltered extension
for d in range(0,self.dim):
#Upper bound
x_boundry=np.max(self.X[d],axis=0)
x_boundry_index=np.argmax(self.X[d],axis=0)
xb=self.X[x_boundry_index]
ucb_y=self.ucb(self.X[x_boundry_index],self.gp)
while(((ucb_y>lcb_y)&(x_boundry<extended_bound[d,1]))|(x_boundry<self.scalebounds[d,1])):
x_boundry=x_boundry+step
xb[d]=xb[d]+step
ucb_y=self.ucb(xb,self.gp)
extended_bound[d,1]=x_boundry
#Lower bound
x_boundry=np.min(self.X[d],axis=0)
ucb_y=self.ucb(self.X[x_boundry_index],self.gp)
while(((ucb_y>lcb_y)&(x_boundry>extended_bound[d,0]))|(x_boundry>self.scalebounds[d,0])):
x_boundry=x_boundry-step
xb[d]=xb[d]-step
ucb_y=self.ucb(xb,self.gp)
extended_bound[d,0]=x_boundry
self.scalebounds=extended_bound
print("seach space extended to {}".format(self.scalebounds))
def volumeDoubling(self):
"""
Description: Expands the search space with the volume doubling method
by Shahriari et al
"""
print('Attempting to expand search space with volume doubling method')
extended_bound=np.copy(self.scalebounds)
old_radius=(extended_bound[:,1]-extended_bound[:,0])/2
volume=np.power(2*old_radius,self.dim)
mid_point=extended_bound[:,0]+old_radius
new_radius=np.power(2*volume,1/self.dim)/2
extended_bound[:,0]=mid_point-new_radius
extended_bound[:,1]=mid_point+new_radius
self.scalebounds=extended_bound
print("seach space extended to {}".format(self.scalebounds))
def maximize(self,gp_params):
"""
Main optimization method.
Input parameters
----------
gp_params: parameter for Gaussian Process
Returns
-------
x: recommented point for evaluation
"""
if self.stop_flag==1:
return
if self.acq['name']=='random':
x_max = [np.random.uniform(x[0], x[1], size=1) for x in self.scalebounds]
x_max=np.asarray(x_max)
x_max=x_max.T
self.X_original=np.vstack((self.X_original, x_max))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y=(self.Y_original-np.mean(self.Y_original))/np.std(self.Y_original)
self.time_opt=np.hstack((self.time_opt,0))
return
# init a new Gaussian Process
self.gp=PradaGaussianProcess(gp_params)
if self.gp.KK_x_x_inv ==[]:
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
acq=self.acq
if acq['debug']==1:
logmarginal=self.gp.log_marginal_lengthscale(gp_params['theta'],gp_params['noise_delta'])
print(gp_params['theta'])
print("log marginal before optimizing ={:.4f}".format(logmarginal))
self.logmarginal=logmarginal
if logmarginal<-999999:
logmarginal=self.gp.log_marginal_lengthscale(gp_params['theta'],gp_params['noise_delta'])
if self.optimize_gp==1 and len(self.Y)%2*self.dim==0 and len(self.Y)>5*self.dim:
print("Initial length scale={}".format(gp_params['theta']))
newtheta = self.gp.optimize_lengthscale(gp_params['theta'],gp_params['noise_delta'],self.scalebounds)
gp_params['theta']=newtheta
print("New length scale={}".format(gp_params['theta']))
# init a new Gaussian Process after optimizing hyper-parameter
self.gp=PradaGaussianProcess(gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Modify search space based on selected method
if self.expandSS=='expandBoundsDDB_MAP':
self.expandBoundsDDB_MAP()
if self.expandSS=='expandBoundsDDB_FB':
self.expandBoundsDDB_FB()
if self.expandSS=='expandBoundsFiltering':
self.expandBoundsFiltering()
if self.expandSS=='volumeDoubling' and len(self.Y)%3*self.dim==0:
self.volumeDoubling()
# Prevent bounds from breaching maximum limit
for d in range(0,self.dim):
if self.scalebounds[d,0]<0:
print('Lower bound of {} in dimention {} exceeded minimum bound of {}. Scaling up.'.format(self.scalebounds[d,0],d,0))
self.scalebounds[d,0]=0
print('bound set to {}'.format(self.scalebounds))
if self.scalebounds[d,1]>max_bound_size:
print('Upper bound of {} in dimention {} exceeded maximum bound of {}. Scaling down.'.format(self.scalebounds[d,1],d,max_bound_size))
self.scalebounds[d,1]=max_bound_size
self.scalebounds[d,0]=min(self.scalebounds[d,0],self.scalebounds[d,1]-np.sqrt(3*self.gp.lengthscale))
print('bound set to {}'.format(self.scalebounds))
# Set acquisition function
start_opt=time.time()
y_max = self.Y.max()
if acq['name'] in ['consensus','mes']:
ucb_acq_func={}
ucb_acq_func['name']='ucb'
ucb_acq_func['kappa']=np.log(len(self.Y))
ucb_acq_func['dim']=self.dim
ucb_acq_func['scalebounds']=self.scalebounds
myacq=AcquisitionFunction(ucb_acq_func)
xt_ucb = acq_max(ac=myacq.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
xstars=[]
xstars.append(xt_ucb)
ei_acq_func={}
ei_acq_func['name']='ei'
ei_acq_func['dim']=self.dim
ei_acq_func['scalebounds']=self.scalebounds
myacq=AcquisitionFunction(ei_acq_func)
xt_ei = acq_max(ac=myacq.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
xstars.append(xt_ei)
pes_acq_func={}
pes_acq_func['name']='pes'
pes_acq_func['dim']=self.dim
pes_acq_func['scalebounds']=self.scalebounds
myacq=AcquisitionFunction(pes_acq_func)
xt_pes = acq_max(ac=myacq.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds)
xstars.append(xt_pes)
self.xstars=xstars
if acq['name']=='vrs':
print("please call the maximize_vrs function")
return
if 'xstars' not in globals():
xstars=[]
self.xstars=xstars
self.acq['xstars']=xstars
self.acq['WW']=False
self.acq['WW_dim']=False
self.acq_func = AcquisitionFunction(self.acq,self.bb_function)
if acq['name']=="ei_mu":
#find the maximum in the predictive mean
mu_acq={}
mu_acq['name']='mu'
mu_acq['dim']=self.dim
acq_mu=AcquisitionFunction(mu_acq)
x_mu_max = acq_max(ac=acq_mu.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds,opt_toolbox=self.opt_toolbox)
# set y_max = mu_max
y_max=acq_mu.acq_kind(x_mu_max,gp=self.gp, y_max=y_max)
x_max = acq_max(ac=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,bounds=self.scalebounds,opt_toolbox=self.opt_toolbox,seeds=self.xstars)
if acq['name']=='consensus' and acq['debug']==1: # plot the x_max and xstars
fig=plt.figure(figsize=(5, 5))
plt.scatter(xt_ucb[0],xt_ucb[1],marker='s',color='g',s=200,label='Peak')
plt.scatter(xt_ei[0],xt_ei[1],marker='s',color='k',s=200,label='Peak')
plt.scatter(x_max[0],x_max[1],marker='*',color='r',s=300,label='Peak')
plt.xlim(0,1)
plt.ylim(0,1)
strFileName="acquisition_functions_debug.eps"
fig.savefig(strFileName, bbox_inches='tight')
if acq['name']=='vrs' and acq['debug']==1: # plot the x_max and xstars
fig=plt.figure(figsize=(5, 5))
plt.scatter(xt_ucb[0],xt_ucb[1],marker='s',color='g',s=200,label='Peak')
plt.scatter(xt_ei[0],xt_ei[1],marker='s',color='k',s=200,label='Peak')
plt.scatter(x_max[0],x_max[1],marker='*',color='r',s=300,label='Peak')
plt.xlim(0,1)
plt.ylim(0,1)
strFileName="vrs_acquisition_functions_debug.eps"
#fig.savefig(strFileName, bbox_inches='tight')
val_acq=self.acq_func.acq_kind(x_max,self.gp,y_max)
#print x_max
#print val_acq
if self.stopping_criteria!=0 and val_acq<self.stopping_criteria:
val_acq=self.acq_func.acq_kind(x_max,self.gp,y_max)
self.stop_flag=1
print("Stopping Criteria is violated. Stopping Criteria is {:.15f}".format(self.stopping_criteria))
self.alpha_Xt= np.append(self.alpha_Xt,val_acq)
mean,var=self.gp.predict(x_max, eval_MSE=True)
var.flags['WRITEABLE']=True
var[var<1e-20]=0
#self.Tau_Xt= np.append(self.Tau_Xt,val_acq/var)
# record the optimization time
finished_opt=time.time()
elapse_opt=finished_opt-start_opt
self.time_opt=np.hstack((self.time_opt,elapse_opt))
# store X
self.X = np.vstack((self.X, x_max.reshape((1, -1))))
# evaluate Y using original X
self.Y_original = np.append(self.Y_original, self.f(x_max))
# update Y after change Y_original
self.Y=(self.Y_original-np.mean(self.Y_original))/np.std(self.Y_original)
if self.gp.flagIncremental==1:
self.gp.fit_incremental(x_max,self.Y[-1])
# if (self.acq['name']=='ei_regularizerH') or (self.acq['name']=='ei_regularizerQ'):
# self.scalebounds[:,0]=self.scalebounds[:,0]+1
# self.scalebounds[:,1]=self.scalebounds[:,1]-1
# self.acq['scalebounds']=self.scalebounds
self.experiment_num=self.experiment_num+1
def maximize(self,
init_points=5,
n_iter=25,
acq='ucb',
kappa=2.576,
**gp_params):
"""
Main optimization method.
Parameters
----------
:param init_points:
Number of randomly chosen points to sample the
target function before fitting the gp.
:param n_iter:
Total number of times the process is to repeated. Note that
currently this methods does not have stopping criteria (due to a
number of reasons), therefore the total number of points to be
sampled must be specified.
:param acq:
Acquisition function to be used, defaults to Expected Improvement.
:param gp_params:
Parameters to be passed to the Scikit-learn Gaussian Process object
Returns
-------
:return: Nothing
"""
# Reset timer
#self.plog.reset_timer()
# Set acquisition function
self.acq_func = AcquisitionFunction(kind=self.acq, kappa=kappa)
# Initialize x, y and find current y_max
if not self.initialized:
#if self.verbose:
#self.plog.print_header()
self.init(init_points)
y_max = self.Y.max()
self.theta = gp_params['theta']
# Set parameters if any was passed
#self.gp.set_params(**gp_params)
self.gp = PradaMultipleGaussianProcess(**gp_params)
# Find unique rows of X to avoid GP from breaking
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
# Finding argmax of the acquisition function.
x_max = acq_max(ac=self.acq_func.acq_kind,
gp=self.gp,
y_max=y_max,
bounds=self.bounds)
#print "start acq max nlopt"
#x_max,f_max = acq_max_nlopt(f=self.acq_func.acq_kind,gp=self.gp,y_max=y_max,
#bounds=self.bounds)
#print "end acq max nlopt"
# Test if x_max is repeated, if it is, draw another one at random
# If it is repeated, print a warning
#pwarning = False
if np.any((self.X - x_max).sum(axis=1) == 0):
#print "x max uniform random"
x_max = np.random.uniform(self.bounds[:, 0],
self.bounds[:, 1],
size=self.bounds.shape[0])
#print "start append X,Y"
self.X = np.vstack((self.X, x_max.reshape((1, -1))))
#self.Y = np.append(self.Y, self.f(**dict(zip(self.keys, x_max))))
self.Y = np.append(self.Y, self.f(x_max))
#print "end append X,Y"
#print 'x_max={:f}'.format(x_max[0])
#print "start fitting GP"
# Updating the GP.
ur = unique_rows(self.X)
self.gp.fit(self.X[ur], self.Y[ur])
#print "end fitting GP"
# Update maximum value to search for next probe point.
if self.Y[-1] > y_max:
y_max = self.Y[-1]